Add files using upload-large-folder tool

analyze_answer_bias.py

@@ -0,0 +1,255 @@
+#!/usr/bin/env python3
+"""
+Answer/Prediction Bias 분석 스크립트
+
+1. GT Answer Distribution: 정답이 A, B, C, D 중 어디에 편중되어 있는지
+2. Model Prediction Distribution: 모델이 특정 선택지를 더 많이 선택하는지
+
+Usage:
+    python experiments/analyze_answer_bias.py <model_result.xlsx> [--subset far_close]
+    python experiments/analyze_answer_bias.py --compare <file1.xlsx> <file2.xlsx> ...
+"""
+
+import argparse
+import sys
+import pandas as pd
+import numpy as np
+from pathlib import Path
+from typing import Dict, List
+from collections import Counter
+
+
+class TeeWriter:
+    """stdout을 터미널과 파일에 동시에 출력"""
+    def __init__(self, filepath):
+        self.terminal = sys.stdout
+        self.file = open(filepath, 'w', encoding='utf-8')
+
+    def write(self, message):
+        self.terminal.write(message)
+        self.file.write(message)
+
+    def flush(self):
+        self.terminal.flush()
+        self.file.flush()
+
+    def close(self):
+        self.file.close()
+        return self.terminal
+
+
+def extract_answer_letter(val) -> str:
+    """예측값에서 A/B/C/D 추출 (예: 'D. basket' -> 'D')"""
+    if pd.isna(val):
+        return 'INVALID'
+    val = str(val).strip()
+    if len(val) == 0:
+        return 'INVALID'
+    first_char = val[0].upper()
+    if first_char in ['A', 'B', 'C', 'D']:
+        return first_char
+    return 'INVALID'
+
+
+def analyze_bias(df: pd.DataFrame, subset_name: str = "ALL") -> Dict:
+    """
+    Answer/Prediction bias 분석
+    """
+    # GT Answer 분포 (이미 A/B/C/D 형태)
+    gt_dist = Counter(df['answer'])
+    gt_total = sum(gt_dist.values())
+
+    # Prediction 분포 - 첫 글자 추출
+    pred_letters = df['prediction'].apply(extract_answer_letter)
+    pred_dist = Counter(pred_letters)
+    pred_total = sum(v for k, v in pred_dist.items() if k != 'INVALID')
+
+    # 정답률 by GT position
+    acc_by_pos = {}
+    for ans in ['A', 'B', 'C', 'D']:
+        subset = df[df['answer'] == ans]
+        if len(subset) > 0:
+            acc_by_pos[ans] = subset['hit'].mean() * 100
+        else:
+            acc_by_pos[ans] = 0
+
+    # Prediction이 GT와 일치하는 비율 (hit rate by prediction position)
+    # prediction에서 첫 글자 추출해서 매칭
+    df_with_pred_letter = df.copy()
+    df_with_pred_letter['pred_letter'] = pred_letters.values
+    hit_by_pred = {}
+    for pred in ['A', 'B', 'C', 'D']:
+        subset = df_with_pred_letter[df_with_pred_letter['pred_letter'] == pred]
+        if len(subset) > 0:
+            hit_by_pred[pred] = subset['hit'].mean() * 100
+        else:
+            hit_by_pred[pred] = 0
+
+    return {
+        'subset': subset_name,
+        'total': len(df),
+        'gt_dist': {k: gt_dist.get(k, 0) for k in ['A', 'B', 'C', 'D']},
+        'gt_pct': {k: gt_dist.get(k, 0) / gt_total * 100 if gt_total > 0 else 0 for k in ['A', 'B', 'C', 'D']},
+        'pred_dist': {k: pred_dist.get(k, 0) for k in ['A', 'B', 'C', 'D']},
+        'pred_pct': {k: pred_dist.get(k, 0) / pred_total * 100 if pred_total > 0 else 0 for k in ['A', 'B', 'C', 'D']},
+        'acc_by_gt_pos': acc_by_pos,
+        'hit_by_pred_pos': hit_by_pred,
+        'overall_acc': df['hit'].mean() * 100
+    }
+
+
+def print_bias_report(xlsx_path: str, results: List[Dict]):
+    """Bias 분석 리포트 출력"""
+    model_name = Path(xlsx_path).stem.replace('_EmbSpatialBench_openai_result', '')
+    # 이름 축약
+    if len(model_name) > 50:
+        model_name = model_name[:47] + "..."
+
+    print(f"\n{'='*80}")
+    print(f"Model: {model_name}")
+    print(f"{'='*80}")
+
+    for r in results:
+        print(f"\n--- {r['subset']} (n={r['total']}) ---")
+
+        # GT Distribution
+        print(f"\n  GT Answer Distribution:")
+        print(f"    {'Pos':<5} {'Count':<8} {'Pct':<8} {'Acc when GT':<12}")
+        print(f"    {'-'*35}")
+        for pos in ['A', 'B', 'C', 'D']:
+            print(f"    {pos:<5} {r['gt_dist'][pos]:<8} {r['gt_pct'][pos]:.1f}%{'':<4} {r['acc_by_gt_pos'][pos]:.1f}%")
+
+        # Prediction Distribution
+        print(f"\n  Model Prediction Distribution:")
+        print(f"    {'Pos':<5} {'Count':<8} {'Pct':<8} {'Acc when Pred':<12}")
+        print(f"    {'-'*35}")
+        for pos in ['A', 'B', 'C', 'D']:
+            print(f"    {pos:<5} {r['pred_dist'][pos]:<8} {r['pred_pct'][pos]:.1f}%{'':<4} {r['hit_by_pred_pos'][pos]:.1f}%")
+
+        # Bias 지표
+        gt_std = np.std([r['gt_pct'][p] for p in ['A', 'B', 'C', 'D']])
+        pred_std = np.std([r['pred_pct'][p] for p in ['A', 'B', 'C', 'D']])
+
+        print(f"\n  Bias Indicators:")
+        print(f"    GT Distribution Std: {gt_std:.2f}%p (uniform=0)")
+        print(f"    Pred Distribution Std: {pred_std:.2f}%p (uniform=0)")
+        print(f"    Overall Accuracy: {r['overall_acc']:.1f}%")
+
+
+def analyze_model(xlsx_path: str, include_subsets: bool = True) -> List[Dict]:
+    """모델 결과 분석"""
+    df = pd.read_excel(xlsx_path)
+
+    results = []
+
+    # 전체 분석
+    results.append(analyze_bias(df, "ALL"))
+
+    if include_subsets:
+        # FAR/CLOSE만 분석
+        far_close_df = df[df['category'].isin(['far', 'close'])]
+        if len(far_close_df) > 0:
+            results.append(analyze_bias(far_close_df, "FAR+CLOSE"))
+
+        # Category별 분석
+        for cat in ['far', 'close']:
+            cat_df = df[df['category'] == cat]
+            if len(cat_df) > 0:
+                results.append(analyze_bias(cat_df, cat.upper()))
+
+    return results
+
+
+def compare_models_bias(xlsx_paths: List[str]):
+    """여러 모델의 bias 비교 (요약 테이블)"""
+
+    print(f"\n{'='*100}")
+    print("MODEL BIAS COMPARISON SUMMARY")
+    print(f"{'='*100}")
+
+    # Header
+    print(f"\n{'Model':<45} {'Subset':<12} {'GT Std':<10} {'Pred Std':<10} {'Pred Max':<12} {'Acc':<8}")
+    print("-" * 97)
+
+    for xlsx_path in xlsx_paths:
+        model_name = Path(xlsx_path).stem.replace('_EmbSpatialBench_openai_result', '')
+        if len(model_name) > 43:
+            model_name = model_name[:40] + "..."
+
+        results = analyze_model(xlsx_path, include_subsets=True)
+
+        for r in results:
+            gt_std = np.std([r['gt_pct'][p] for p in ['A', 'B', 'C', 'D']])
+            pred_std = np.std([r['pred_pct'][p] for p in ['A', 'B', 'C', 'D']])
+
+            # 가장 많이 선택한 position
+            max_pred_pos = max(r['pred_pct'], key=r['pred_pct'].get)
+            max_pred_pct = r['pred_pct'][max_pred_pos]
+
+            if r['subset'] == 'ALL':
+                print(f"{model_name:<45} {r['subset']:<12} {gt_std:.1f}%p{'':<4} {pred_std:.1f}%p{'':<4} {max_pred_pos}({max_pred_pct:.1f}%){'':<2} {r['overall_acc']:.1f}%")
+            else:
+                print(f"{'':<45} {r['subset']:<12} {gt_std:.1f}%p{'':<4} {pred_std:.1f}%p{'':<4} {max_pred_pos}({max_pred_pct:.1f}%){'':<2} {r['overall_acc']:.1f}%")
+
+
+EVAL_OUTPUT_DIR = 'VLMEvalKit/outputs'
+
+DEFAULT_MODELS = [
+    # Molmo-7B
+    'molmo-7B-O-0924/molmo-7B-O-0924',
+    'molmo-7B-O-0924-data_scale_exp_80k/molmo-7B-O-0924-data_scale_exp_80k',
+    'molmo-7B-O-0924-data_scale_exp_400k/molmo-7B-O-0924-data_scale_exp_400k',
+    'molmo-7B-O-0924-data_scale_exp_800k/molmo-7B-O-0924-data_scale_exp_800k',
+    'molmo-7B-O-0924-data_scale_exp_2m/molmo-7B-O-0924-data_scale_exp_2m',
+    # NVILA-Lite-2B
+    'NVILA-Lite-2B/NVILA-Lite-2B',
+    'NVILA-Lite-2B-data-scale-exp-80k/NVILA-Lite-2B-data-scale-exp-80k',
+    'NVILA-Lite-2B-data-scale-exp-400k/NVILA-Lite-2B-data-scale-exp-400k',
+    'NVILA-Lite-2B-data-scale-exp-800k/NVILA-Lite-2B-data-scale-exp-800k',
+    'NVILA-Lite-2B-data-scale-exp-2m/NVILA-Lite-2B-data-scale-exp-2m',
+    'RoboRefer-2B-SFT/RoboRefer-2B-SFT',
+    # Qwen2.5-VL-3B
+    'Qwen2.5-VL-3B-Instruct/Qwen2.5-VL-3B-Instruct',
+    'Qwen2.5-VL-3B-Instruct-data_scale_exp_80k/Qwen2.5-VL-3B-Instruct-data_scale_exp_80k',
+    'Qwen2.5-VL-3B-Instruct-data_scale_exp_400k/Qwen2.5-VL-3B-Instruct-data_scale_exp_400k',
+    'Qwen2.5-VL-3B-Instruct-data_scale_exp_800k/Qwen2.5-VL-3B-Instruct-data_scale_exp_800k',
+    'Qwen2.5-VL-3B-Instruct-data_scale_exp_2m/Qwen2.5-VL-3B-Instruct-data_scale_exp_2m',
+]
+
+
+def get_default_xlsx_paths():
+    return [f'{EVAL_OUTPUT_DIR}/{m}_EmbSpatialBench_openai_result.xlsx' for m in DEFAULT_MODELS]
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Answer/Prediction Bias 분석')
+    parser.add_argument('xlsx_files', nargs='*', help='Model result xlsx files (없으면 기본 모델 사용)')
+    parser.add_argument('--compare', action='store_true', help='Compare multiple models (summary only)')
+    parser.add_argument('--detail', action='store_true', help='Show detailed report for each model')
+    parser.add_argument('--output', '-o', type=str, help='Save results to file')
+
+    args = parser.parse_args()
+
+    xlsx_files = args.xlsx_files if args.xlsx_files else get_default_xlsx_paths()
+
+    if args.output:
+        tee = TeeWriter(args.output)
+        sys.stdout = tee
+
+    if args.compare and not args.detail:
+        compare_models_bias(xlsx_files)
+    else:
+        for xlsx_path in xlsx_files:
+            results = analyze_model(xlsx_path)
+            print_bias_report(xlsx_path, results)
+
+        if len(xlsx_files) > 1:
+            compare_models_bias(xlsx_files)
+
+    if args.output:
+        sys.stdout = tee.close()
+        print(f"Results saved to {args.output}")
+
+
+if __name__ == '__main__':
+    main()

analyze_counter_consistent.py

@@ -0,0 +1,627 @@
+#!/usr/bin/env python3
+"""
+Counter vs Consistent Example Analysis Script
+
+2D Heuristic (shared across datasets):
+  Upper part of image (small y) = farther from camera
+  Lower part of image (large y) = closer to camera
+
+Datasets:
+  embspatial (default):
+    FAR/CLOSE questions in EmbSpatial-Bench
+    Consistent: GT answer agrees with the 2D heuristic (Height-Depth Entanglement)
+    Counter: GT answer contradicts the 2D heuristic
+
+  cvbench3d:
+    Depth questions: "Which object is closer to the camera?"
+      Consistent: GT object (closer) has larger center_y (lower in image)
+      Counter: GT object (closer) has smaller center_y (higher in image)
+    Distance questions: "Which object is closer to [reference]?"
+      2D heuristic: smaller pixel distance to reference = closer in 3D
+      Consistent: GT candidate has smaller 2D pixel distance to reference
+      Counter: GT candidate has larger 2D pixel distance to reference
+
+Usage:
+    python experiments/analyze_counter_consistent.py <model_result.xlsx> [--verbose]
+    python experiments/analyze_counter_consistent.py --compare <file1.xlsx> <file2.xlsx> ...
+    python experiments/analyze_counter_consistent.py --dataset cvbench3d <result.xlsx>
+    python experiments/analyze_counter_consistent.py --dataset cvbench3d --compare <file1.xlsx> ...
+"""
+
+import argparse
+import ast
+import pandas as pd
+import numpy as np
+from datasets import load_dataset
+from pathlib import Path
+from typing import Dict, List, Tuple, Optional
+import json
+import sys
+
+
+class TeeWriter:
+    """Write stdout to both terminal and file simultaneously"""
+    def __init__(self, filepath):
+        self.terminal = sys.stdout
+        self.file = open(filepath, 'w', encoding='utf-8')
+
+    def write(self, message):
+        self.terminal.write(message)
+        self.file.write(message)
+        self.file.flush()
+
+    def flush(self):
+        self.terminal.flush()
+        self.file.flush()
+
+    def close(self):
+        self.file.close()
+        return self.terminal
+
+
+# =============================================================================
+# EmbSpatial-Bench
+# =============================================================================
+
+def get_bbox_center_y(bbox: List[int], source: str = None) -> float:
+    """
+    BBox -> center y coordinate, format varies by source:
+      ScanNet / MP3D : [x1, y1, w,  h ] -> y1 + h/2
+      AI2Thor        : [x1, y1, x2, y2] -> (y1 + y2) / 2
+    """
+    if source == 'ai2thor':
+        return (bbox[1] + bbox[3]) / 2
+    else:
+        return bbox[1] + bbox[3] / 2
+
+
+def classify_sample(relation: str, objects: Dict, gt_answer_idx: int,
+                    answer_options: List[str] = None,
+                    image_height: int = None, threshold_ratio: float = 0.05,
+                    data_source: str = None) -> Tuple[str, Dict]:
+    """
+    Classify a sample as Consistent / Counter / Ambiguous.
+
+    Args:
+        relation: 'far' or 'close'
+        objects: {'bbox': [...], 'name': [...]}
+        gt_answer_idx: GT answer index (0-based, relative to answer_options)
+        answer_options: list of answer choices (used to match bbox by name)
+        image_height: image height for threshold normalization (pass PIL image.size[1])
+        threshold_ratio: ambiguous decision threshold as a fraction of image height
+        data_source: 'scannet' | 'mp3d' | 'ai2thor' (selects bbox format)
+
+    Returns:
+        classification: 'consistent', 'counter', or 'ambiguous'
+        details: dict with classification details
+    """
+    if relation not in ['far', 'close']:
+        return 'not_applicable', {}
+
+    bboxes = objects['bbox']
+    names = objects['name']
+
+    if len(bboxes) < 2:
+        return 'insufficient_objects', {}
+
+    # answer_options and objects['name'] may differ (e.g. 'Unknown')
+    # resolve GT answer index against objects['name']
+    if answer_options is not None and gt_answer_idx < len(answer_options):
+        gt_answer_name = answer_options[gt_answer_idx]
+        if gt_answer_name in names:
+            gt_answer_idx = names.index(gt_answer_name)
+        elif gt_answer_name == 'Unknown' or gt_answer_idx >= len(bboxes):
+            return 'unknown_object', {}
+
+    # bounds check
+    if gt_answer_idx >= len(bboxes):
+        return 'index_out_of_range', {}
+
+    # compute center y per object using source-specific bbox format
+    center_ys = [get_bbox_center_y(bbox, source=data_source) for bbox in bboxes]
+
+    # GT object center y
+    gt_center_y = center_ys[gt_answer_idx]
+
+    # mean center y of all other objects
+    other_ys = [y for i, y in enumerate(center_ys) if i != gt_answer_idx]
+    other_avg_y = np.mean(other_ys)
+
+    # y difference
+    y_diff = gt_center_y - other_avg_y
+
+    # threshold normalized by image height
+    if image_height:
+        threshold = image_height * threshold_ratio
+    else:
+        threshold = 20  # fallback: 20 pixels
+
+    details = {
+        'gt_object': names[gt_answer_idx],
+        'gt_center_y': gt_center_y,
+        'other_avg_y': other_avg_y,
+        'y_diff': y_diff,
+        'threshold': threshold,
+        'all_objects': list(zip(names, center_ys))
+    }
+
+    # ambiguous if difference is too small
+    if abs(y_diff) < threshold:
+        return 'ambiguous', details
+
+    # FAR: consistent if GT is higher (smaller y)
+    if relation == 'far':
+        if gt_center_y < other_avg_y:
+            return 'consistent', details
+        else:
+            return 'counter', details
+
+    # CLOSE: consistent if GT is lower (larger y)
+    else:
+        if gt_center_y > other_avg_y:
+            return 'consistent', details
+        else:
+            return 'counter', details
+
+
+def get_image_height_by_source(data_source: str) -> int:
+    """Return fallback image height by data source (used when PIL image is unavailable)"""
+    heights = {
+        'ai2thor': 300,
+        'mp3d': 480,
+        'scannet': 968,
+    }
+    return heights.get(data_source, 480)
+
+
+def build_classification_cache(verbose: bool = False) -> Dict[str, Dict]:
+    """
+    Build a counter/consistent classification cache for the full EmbSpatial-Bench dataset.
+    """
+    print("Loading EmbSpatial-Bench dataset...")
+    ds = load_dataset('FlagEval/EmbSpatial-Bench', split='test')
+
+    cache = {}
+    stats = {'far': {'consistent': 0, 'counter': 0, 'ambiguous': 0},
+             'close': {'consistent': 0, 'counter': 0, 'ambiguous': 0}}
+
+    for item in ds:
+        question_id = item['question_id']
+        relation = item['relation']
+
+        if relation not in ['far', 'close']:
+            cache[question_id] = {'classification': 'not_applicable', 'relation': relation}
+            continue
+
+        objects = item['objects']
+        gt_answer_idx = item['answer']  # 0-based index
+        answer_options = item['answer_options']
+        data_source = item['data_source']
+
+        # use actual image height from PIL image (image.size -> (width, height))
+        pil_image = item.get('image')
+        if pil_image is not None and hasattr(pil_image, 'size'):
+            image_height = pil_image.size[1]
+        else:
+            image_height = get_image_height_by_source(data_source)
+
+        classification, details = classify_sample(
+            relation, objects, gt_answer_idx, answer_options, image_height,
+            data_source=data_source
+        )
+
+        cache[question_id] = {
+            'classification': classification,
+            'relation': relation,
+            'data_source': item['data_source'],
+            'details': details
+        }
+
+        if relation in stats and classification in stats[relation]:
+            stats[relation][classification] += 1
+
+    if verbose:
+        print("\n=== Classification Statistics ===")
+        for rel in ['far', 'close']:
+            total = sum(stats[rel].values())
+            print(f"\n{rel.upper()} (n={total}):")
+            for cls, cnt in stats[rel].items():
+                pct = cnt / total * 100 if total > 0 else 0
+                print(f"  {cls}: {cnt} ({pct:.1f}%)")
+
+    return cache
+
+
+def analyze_embspatial_results(xlsx_path: str, cache: Dict[str, Dict],
+                               verbose: bool = False) -> Tuple[Dict, List[Dict]]:
+    """Analyze a model result xlsx file against the EmbSpatialBench classification cache."""
+    df = pd.read_excel(xlsx_path)
+
+    results = {
+        'far': {
+            'consistent': {'correct': 0, 'total': 0},
+            'counter': {'correct': 0, 'total': 0},
+            'ambiguous': {'correct': 0, 'total': 0}
+        },
+        'close': {
+            'consistent': {'correct': 0, 'total': 0},
+            'counter': {'correct': 0, 'total': 0},
+            'ambiguous': {'correct': 0, 'total': 0}
+        }
+    }
+
+    counter_examples = []
+
+    for _, row in df.iterrows():
+        question_id = row['question_id']
+        category = row['category']
+        hit = row['hit']
+
+        if category not in ['far', 'close']:
+            continue
+
+        if question_id not in cache:
+            continue
+
+        info = cache[question_id]
+        classification = info['classification']
+
+        if classification not in ['consistent', 'counter', 'ambiguous']:
+            continue
+
+        results[category][classification]['total'] += 1
+        if hit == 1:
+            results[category][classification]['correct'] += 1
+
+        if classification == 'counter':
+            counter_examples.append({
+                'question_id': question_id,
+                'relation': category,
+                'hit': hit,
+                'prediction': row['prediction'],
+                'answer': row['answer'],
+                'data_source': info['data_source'],
+                'details': info.get('details', {})
+            })
+
+    return results, counter_examples
+
+
+# =============================================================================
+# CV-Bench-3D
+# =============================================================================
+
+# Known image heights per source dataset (used for threshold normalization)
+# Omni3D_SUNRGBD has variable sizes; fallback to max bbox y2 estimate.
+_CVBENCH3D_SOURCE_HEIGHTS = {
+    'Omni3D_Hypersim': 768,
+    'Omni3D_nuScenes': 900,
+}
+
+
+def classify_cvbench3d_row(row, depth_threshold_ratio: float = 0.05) -> Tuple[str, Dict]:
+    """
+    Classify a single CV-Bench-3D row as consistent / counter / ambiguous.
+
+    Only Depth questions are classified — they share the same height-depth
+    entanglement heuristic as EmbSpatial-Bench:
+        2D heuristic: lower in image (larger center_y) = closer to camera
+        Consistent: GT object (closer to camera) has larger center_y
+        Counter:     GT object (closer to camera) has smaller center_y
+
+    Distance questions ask "which object is closer to [reference] in 3D real-world
+    distance?" — this is inter-object 3D distance, not viewer distance.  No
+    equivalent 2D projection heuristic exists (height-depth entanglement does not
+    apply), so Distance rows are always marked 'not_applicable'.
+    """
+    category = row['category']
+    answer_letter = str(row['answer']).strip()
+
+    if category != 'Depth':
+        return 'not_applicable', {}
+
+    try:
+        bbox_list = ast.literal_eval(row['bbox'])
+    except (ValueError, SyntaxError):
+        return 'invalid_bbox', {}
+
+    if len(bbox_list) != 2:
+        return 'invalid_bbox', {}
+
+    cy_A = (bbox_list[0][1] + bbox_list[0][3]) / 2
+    cy_B = (bbox_list[1][1] + bbox_list[1][3]) / 2
+
+    gt_y = cy_A if answer_letter == 'A' else cy_B
+    other_y = cy_B if answer_letter == 'A' else cy_A
+    y_diff = gt_y - other_y  # positive = GT is lower in image
+
+    # Estimate image height: prefer known source height, fall back to max bbox y2
+    source_dataset = str(row.get('source_dataset', ''))
+    known_h = _CVBENCH3D_SOURCE_HEIGHTS.get(source_dataset, 0)
+    est_h = max(bb[3] for bb in bbox_list)
+    image_height = max(known_h, est_h)
+    threshold = image_height * depth_threshold_ratio
+
+    details = {
+        'answer': answer_letter,
+        'center_y_A': cy_A,
+        'center_y_B': cy_B,
+        'y_diff': y_diff,
+        'threshold': threshold,
+        'image_height_est': image_height,
+        'source_dataset': source_dataset,
+    }
+
+    if abs(y_diff) < threshold:
+        return 'ambiguous', details
+    # Consistent: GT (closer to camera) is lower in image (larger y)
+    return ('consistent' if gt_y > other_y else 'counter'), details
+
+
+def analyze_cvbench3d_results(xlsx_path: str, verbose: bool = False,
+                              depth_threshold_ratio: float = 0.05) -> Tuple[Dict, List[Dict]]:
+    """
+    Analyze a CV-Bench-3D result xlsx file.
+
+    Only the Depth category is classified into consistent / counter / ambiguous,
+    because it shares the height-depth entanglement heuristic with EmbSpatial-Bench.
+    Distance (inter-object 3D distance) has no analogous 2D projection heuristic
+    and is excluded from the consistent/counter analysis.
+    """
+    df = pd.read_excel(xlsx_path)
+
+    results = {
+        'Depth': {
+            'consistent': {'correct': 0, 'total': 0},
+            'counter': {'correct': 0, 'total': 0},
+            'ambiguous': {'correct': 0, 'total': 0},
+        },
+        # Distance: excluded — no height-depth entanglement heuristic for inter-object distance
+    }
+
+    counter_examples = []
+
+    for _, row in df.iterrows():
+        category = row['category']
+        if category != 'Depth':
+            continue
+
+        hit = row['hit']
+        classification, details = classify_cvbench3d_row(row, depth_threshold_ratio)
+
+        if classification not in ['consistent', 'counter', 'ambiguous']:
+            continue
+
+        results['Depth'][classification]['total'] += 1
+        if hit == 1:
+            results['Depth'][classification]['correct'] += 1
+
+        if classification == 'counter':
+            counter_examples.append({
+                'index': row['index'],
+                'category': category,
+                'hit': hit,
+                'prediction': row['prediction'],
+                'answer': row['answer'],
+                'source_dataset': row.get('source_dataset', ''),
+                'details': details,
+            })
+
+    if verbose:
+        print("\n=== CV-Bench-3D Depth Classification Statistics ===")
+        total = sum(results['Depth'][c]['total'] for c in ['consistent', 'counter', 'ambiguous'])
+        print(f"Depth (n={total}):")
+        for cls in ['consistent', 'counter', 'ambiguous']:
+            n = results['Depth'][cls]['total']
+            pct = n / total * 100 if total > 0 else 0
+            print(f"  {cls}: {n} ({pct:.1f}%)")
+        print("(Distance excluded: no 2D heuristic applies for inter-object 3D distance)")
+
+    return results, counter_examples
+
+
+# =============================================================================
+# Generic report / compare (works for both datasets)
+# =============================================================================
+
+_XLSX_SUFFIXES = {
+    'embspatial': [
+        '_EmbSpatialBench_openai_result',
+        '_EmbSpatialBench_exact_matching_result',
+    ],
+    'cvbench3d': [
+        '_CV-Bench-3D_chatgpt-0125_result',
+        '_CV-Bench-3D_exact_matching_result',
+    ],
+}
+
+
+def extract_model_name(xlsx_path: str, dataset: str) -> str:
+    stem = Path(xlsx_path).stem
+    for suffix in _XLSX_SUFFIXES.get(dataset, []):
+        stem = stem.replace(suffix, '')
+    return stem
+
+
+def print_analysis_report(xlsx_path: str, results: Dict, counter_examples: List[Dict],
+                          dataset: str) -> Dict:
+    """Print analysis report for a single model (works for any dataset)."""
+    model_name = extract_model_name(xlsx_path, dataset)
+
+    print(f"\n{'='*70}")
+    print(f"Model: {model_name}")
+    print(f"{'='*70}")
+
+    print(f"\n{'Category':<12} {'Type':<12} {'Correct':<10} {'Total':<10} {'Accuracy':<10}")
+    print("-" * 54)
+
+    total_consistent = {'correct': 0, 'total': 0}
+    total_counter = {'correct': 0, 'total': 0}
+
+    for category in results:
+        for cls_type in ['consistent', 'counter', 'ambiguous']:
+            data = results[category][cls_type]
+            if data['total'] > 0:
+                acc = data['correct'] / data['total'] * 100
+                print(f"{category:<12} {cls_type:<12} {data['correct']:<10} {data['total']:<10} {acc:.1f}%")
+
+                if cls_type == 'consistent':
+                    total_consistent['correct'] += data['correct']
+                    total_consistent['total'] += data['total']
+                elif cls_type == 'counter':
+                    total_counter['correct'] += data['correct']
+                    total_counter['total'] += data['total']
+
+    print("-" * 54)
+    if total_consistent['total'] > 0:
+        acc = total_consistent['correct'] / total_consistent['total'] * 100
+        print(f"{'TOTAL':<12} {'consistent':<12} {total_consistent['correct']:<10} {total_consistent['total']:<10} {acc:.1f}%")
+    if total_counter['total'] > 0:
+        acc = total_counter['correct'] / total_counter['total'] * 100
+        print(f"{'TOTAL':<12} {'counter':<12} {total_counter['correct']:<10} {total_counter['total']:<10} {acc:.1f}%")
+
+    if total_consistent['total'] > 0 and total_counter['total'] > 0:
+        consistent_acc = total_consistent['correct'] / total_consistent['total'] * 100
+        counter_acc = total_counter['correct'] / total_counter['total'] * 100
+        gap = consistent_acc - counter_acc
+        print(f"\nAccuracy Gap (Consistent - Counter): {gap:.1f}%p")
+        print(f"   -> Larger gap indicates stronger reliance on the 2D heuristic")
+
+    counter_wrong = [ex for ex in counter_examples if ex['hit'] == 0]
+    if len(counter_wrong) > 0:
+        print(f"\n🔍 Counter examples wrong: {len(counter_wrong)} / {len(counter_examples)}")
+
+    return {
+        'model_name': model_name,
+        'consistent_acc': total_consistent['correct'] / total_consistent['total'] * 100 if total_consistent['total'] > 0 else 0,
+        'counter_acc': total_counter['correct'] / total_counter['total'] * 100 if total_counter['total'] > 0 else 0,
+        'consistent_total': total_consistent['total'],
+        'counter_total': total_counter['total'],
+    }
+
+
+def _run_analysis(xlsx_path: str, dataset: str, cache: Optional[Dict] = None,
+                  verbose: bool = False,
+                  depth_threshold_ratio: float = 0.05) -> Tuple[Dict, List[Dict]]:
+    if dataset == 'cvbench3d':
+        return analyze_cvbench3d_results(xlsx_path, verbose=verbose,
+                                         depth_threshold_ratio=depth_threshold_ratio)
+    else:
+        return analyze_embspatial_results(xlsx_path, cache, verbose=verbose)
+
+
+def compare_models(xlsx_paths: List[str], dataset: str, cache: Optional[Dict] = None):
+    """Compare multiple models side by side."""
+    summaries = []
+
+    for xlsx_path in xlsx_paths:
+        results, counter_examples = _run_analysis(xlsx_path, dataset, cache)
+        summary = print_analysis_report(xlsx_path, results, counter_examples, dataset)
+        summaries.append(summary)
+
+    max_name_len = max(len(s['model_name']) for s in summaries)
+    col_w = max(max_name_len + 2, 40)
+    total_w = col_w + 12 + 12 + 10
+    print(f"\n{'='*total_w}")
+    print("MODEL COMPARISON")
+    print(f"{'='*total_w}")
+    print(f"{'Model':<{col_w}} {'Consistent':<12} {'Counter':<12} {'Gap':<10}")
+    print("-" * total_w)
+
+    for s in summaries:
+        gap = s['consistent_acc'] - s['counter_acc']
+        print(f"{s['model_name']:<{col_w}} {s['consistent_acc']:.1f}%{'':<6} {s['counter_acc']:.1f}%{'':<6} {gap:+.1f}%p")
+
+
+EVAL_OUTPUT_DIR = 'VLMEvalKit/outputs'
+
+DEFAULT_MODELS = [
+    # Molmo-7B
+    'molmo-7B-O-0924/molmo-7B-O-0924',
+    'molmo-7B-O-0924-data_scale_exp_80k/molmo-7B-O-0924-data_scale_exp_80k',
+    'molmo-7B-O-0924-data_scale_exp_400k/molmo-7B-O-0924-data_scale_exp_400k',
+    'molmo-7B-O-0924-data_scale_exp_800k/molmo-7B-O-0924-data_scale_exp_800k',
+    'molmo-7B-O-0924-data_scale_exp_2m/molmo-7B-O-0924-data_scale_exp_2m',
+    # NVILA-Lite-2B
+    'NVILA-Lite-2B/NVILA-Lite-2B',
+    'NVILA-Lite-2B-data-scale-exp-80k/NVILA-Lite-2B-data-scale-exp-80k',
+    'NVILA-Lite-2B-data-scale-exp-400k/NVILA-Lite-2B-data-scale-exp-400k',
+    'NVILA-Lite-2B-data-scale-exp-800k/NVILA-Lite-2B-data-scale-exp-800k',
+    'NVILA-Lite-2B-data-scale-exp-2m/NVILA-Lite-2B-data-scale-exp-2m',
+    'NVILA-Lite-2B-ST-80k-5pct/NVILA-Lite-2B-ST-80k-5pct',
+    'NVILA-Lite-2B-ST-400k-5pct/NVILA-Lite-2B-ST-400k-5pct',
+    'NVILA-Lite-2B-ST-800k-5pct/NVILA-Lite-2B-ST-800k-5pct',
+    'RoboRefer-2B-SFT/RoboRefer-2B-SFT',
+    # Qwen2.5-VL-3B
+    'Qwen2.5-VL-3B-Instruct/Qwen2.5-VL-3B-Instruct',
+    'Qwen2.5-VL-3B-Instruct-data_scale_exp_80k/Qwen2.5-VL-3B-Instruct-data_scale_exp_80k',
+    'Qwen2.5-VL-3B-Instruct-data_scale_exp_400k/Qwen2.5-VL-3B-Instruct-data_scale_exp_400k',
+    'Qwen2.5-VL-3B-Instruct-data_scale_exp_800k/Qwen2.5-VL-3B-Instruct-data_scale_exp_800k',
+    'Qwen2.5-VL-3B-Instruct-data_scale_exp_2m/Qwen2.5-VL-3B-Instruct-data_scale_exp_2m',
+    'Qwen3-VL-235B-A22B-Instruct/Qwen3-VL-235B-A22B-Instruct'
+]
+
+
+def get_default_xlsx_paths(dataset: str) -> List[str]:
+    if dataset == 'cvbench3d':
+        return [f'{EVAL_OUTPUT_DIR}/{m}_CV-Bench-3D_chatgpt-0125_result.xlsx'
+                for m in DEFAULT_MODELS]
+    else:
+        return [f'{EVAL_OUTPUT_DIR}/{m}_EmbSpatialBench_openai_result.xlsx'
+                for m in DEFAULT_MODELS]
+
+
+def main():
+    parser = argparse.ArgumentParser(description='Counter vs Consistent Example Analysis')
+    parser.add_argument('xlsx_files', nargs='*',
+                        help='Model result xlsx files (uses default model list if omitted)')
+    parser.add_argument('--dataset', choices=['embspatial', 'cvbench3d'], default='embspatial',
+                        help='Benchmark dataset to analyze (default: embspatial)')
+    parser.add_argument('--compare', action='store_true', help='Compare multiple models')
+    parser.add_argument('--verbose', '-v', action='store_true', help='Verbose output')
+    parser.add_argument('--output', '-o', type=str, help='Save results to file')
+    parser.add_argument('--save-cache', type=str,
+                        help='Save EmbSpatialBench classification cache to JSON')
+    parser.add_argument('--load-cache', type=str,
+                        help='Load EmbSpatialBench classification cache from JSON')
+
+    args = parser.parse_args()
+
+    # Build/load cache (EmbSpatialBench only; CV-Bench-3D reads bbox from xlsx directly)
+    cache = None
+    if args.dataset == 'embspatial':
+        if args.load_cache and Path(args.load_cache).exists():
+            print(f"Loading cache from {args.load_cache}...")
+            with open(args.load_cache, 'r') as f:
+                cache = json.load(f)
+        else:
+            cache = build_classification_cache(verbose=args.verbose)
+
+        if args.save_cache:
+            print(f"Saving cache to {args.save_cache}...")
+            with open(args.save_cache, 'w') as f:
+                json.dump(cache, f, indent=2)
+
+    xlsx_files = args.xlsx_files if args.xlsx_files else get_default_xlsx_paths(args.dataset)
+
+    tee = None
+    if args.output:
+        tee = TeeWriter(args.output)
+        sys.stdout = tee
+
+    try:
+        if args.compare or len(xlsx_files) > 1:
+            compare_models(xlsx_files, args.dataset, cache)
+        else:
+            results, counter_examples = _run_analysis(
+                xlsx_files[0], args.dataset, cache, args.verbose
+            )
+            print_analysis_report(xlsx_files[0], results, counter_examples, args.dataset)
+    finally:
+        if tee is not None:
+            sys.stdout = tee.close()
+            print(f"Results saved to {args.output}")
+
+
+if __name__ == '__main__':
+    main()

analyze_heuristic_position.py

@@ -0,0 +1,256 @@
+#!/usr/bin/env python3
+"""
+2D Heuristic 답의 선택지 위치(A/B/C/D) 분포 분석
+
+가설: FAR 질문에서 2D heuristic 답(이미지 위쪽 = 가장 먼 물체)이
+특정 선택지 위치(예: D)에 편중되어 있으면, D bias를 가진 모델이
+FAR에서 더 강한 bias를 보이는 이유를 설명할 수 있음.
+
+2D Heuristic 답 정의:
+- FAR: center_y가 가장 작은 물체 (이미지 위쪽 = "가장 멀다")
+- CLOSE: center_y가 가장 큰 물체 (이미지 아래쪽 = "가장 가깝다")
+
+Usage:
+    python experiments/analyze_heuristic_position.py
+    python experiments/analyze_heuristic_position.py -o experiments/heuristic_position_results.txt
+"""
+
+import argparse
+import numpy as np
+from datasets import load_dataset
+from collections import Counter, defaultdict
+import sys
+
+
+class TeeWriter:
+    """stdout을 터미널과 파일에 동시에 출력"""
+    def __init__(self, filepath):
+        self.terminal = sys.stdout
+        self.file = open(filepath, 'w', encoding='utf-8')
+
+    def write(self, message):
+        self.terminal.write(message)
+        self.file.write(message)
+
+    def flush(self):
+        self.terminal.flush()
+        self.file.flush()
+
+    def close(self):
+        self.file.close()
+        return self.terminal
+
+
+def get_bbox_center_y(bbox):
+    """BBox [x, y, width, height] -> center y coordinate"""
+    return bbox[1] + bbox[3] / 2
+
+
+def find_heuristic_answer(relation, objects, answer_options):
+    """
+    2D heuristic이 선택할 답을 찾는다.
+
+    FAR: center_y가 가장 작은 물체 (이미지 위쪽 = "가장 멀다")
+    CLOSE: center_y가 가장 큰 물체 (이미지 아래쪽 = "가장 가깝다")
+
+    Returns:
+        heuristic_position: 0~3 (A~D), or None if not found
+        heuristic_name: 물체 이름
+    """
+    bboxes = objects['bbox']
+    names = objects['name']
+
+    if len(bboxes) < 2:
+        return None, None
+
+    # 각 물체의 center_y 계산
+    center_ys = [(i, names[i], get_bbox_center_y(bboxes[i])) for i in range(len(bboxes))]
+
+    if relation == 'far':
+        # 가장 작은 center_y (이미지 위쪽) = heuristic이 "가장 멀다"고 판단
+        heuristic_obj = min(center_ys, key=lambda x: x[2])
+    else:  # close
+        # 가장 큰 center_y (이미지 아래쪽) = heuristic이 "가장 가깝다"고 판단
+        heuristic_obj = max(center_ys, key=lambda x: x[2])
+
+    heuristic_name = heuristic_obj[1]
+
+    # answer_options에서 이 물체의 위치(A/B/C/D) 찾기
+    if heuristic_name in answer_options:
+        position = answer_options.index(heuristic_name)
+        return position, heuristic_name
+
+    return None, heuristic_name
+
+
+def main():
+    parser = argparse.ArgumentParser(description='2D Heuristic 답의 선택지 위치 분포 분석')
+    parser.add_argument('-o', '--output', type=str, help='Save results to file')
+    args = parser.parse_args()
+
+    if args.output:
+        tee = TeeWriter(args.output)
+        sys.stdout = tee
+
+    print("Loading EmbSpatial-Bench dataset...")
+    ds = load_dataset('FlagEval/EmbSpatial-Bench', split='test')
+
+    position_labels = ['A', 'B', 'C', 'D']
+
+    # 전체 통계
+    heuristic_pos_far = []       # FAR에서 heuristic 답의 위치
+    heuristic_pos_close = []     # CLOSE에서 heuristic 답의 위치
+    gt_pos_far = []              # FAR에서 GT 답의 위치
+    gt_pos_close = []            # CLOSE에서 GT 답의 위치
+
+    # heuristic == GT인지 여부
+    heuristic_is_gt_far = 0
+    heuristic_is_gt_close = 0
+    total_far = 0
+    total_close = 0
+
+    # 상세 분석용
+    not_found_count = 0
+
+    for item in ds:
+        relation = item['relation']
+        if relation not in ['far', 'close']:
+            continue
+
+        objects = item['objects']
+        answer_options = item['answer_options']
+        gt_answer_idx = item['answer']
+
+        heuristic_pos, heuristic_name = find_heuristic_answer(relation, objects, answer_options)
+
+        if heuristic_pos is None:
+            not_found_count += 1
+            continue
+
+        if relation == 'far':
+            total_far += 1
+            heuristic_pos_far.append(heuristic_pos)
+            gt_pos_far.append(gt_answer_idx)
+            if heuristic_pos == gt_answer_idx:
+                heuristic_is_gt_far += 1
+        else:
+            total_close += 1
+            heuristic_pos_close.append(heuristic_pos)
+            gt_pos_close.append(gt_answer_idx)
+            if heuristic_pos == gt_answer_idx:
+                heuristic_is_gt_close += 1
+
+    # ===== 결과 출력 =====
+    print(f"\n{'='*70}")
+    print("2D Heuristic 답의 선택지 위치(A/B/C/D) 분포 분석")
+    print(f"{'='*70}")
+    print(f"\nHeuristic 정의:")
+    print(f"  FAR:   center_y가 가장 작은 물체 (이미지 위쪽 = '가장 멀다')")
+    print(f"  CLOSE: center_y가 가장 큰 물체 (이미지 아래쪽 = '가장 가깝다')")
+    print(f"\n매칭 실패: {not_found_count}개 (answer_options에 heuristic 물체가 없음)")
+
+    for label, h_positions, g_positions, total, h_is_gt in [
+        ('FAR', heuristic_pos_far, gt_pos_far, total_far, heuristic_is_gt_far),
+        ('CLOSE', heuristic_pos_close, gt_pos_close, total_close, heuristic_is_gt_close),
+        ('FAR+CLOSE', heuristic_pos_far + heuristic_pos_close,
+         gt_pos_far + gt_pos_close, total_far + total_close,
+         heuristic_is_gt_far + heuristic_is_gt_close),
+    ]:
+        print(f"\n{'─'*60}")
+        print(f"  {label} (n={total})")
+        print(f"{'─'*60}")
+
+        # Heuristic 답 위치 분포
+        h_counter = Counter(h_positions)
+        print(f"\n  [Heuristic 답의 위치 분포]")
+        print(f"  {'Position':<10} {'Count':<10} {'Ratio':<10}")
+        for i, pl in enumerate(position_labels):
+            cnt = h_counter.get(i, 0)
+            ratio = cnt / total * 100 if total > 0 else 0
+            print(f"  {pl:<10} {cnt:<10} {ratio:.1f}%")
+        h_std = np.std([h_counter.get(i, 0) / total * 100 for i in range(4)])
+        print(f"  Std: {h_std:.1f}%p")
+
+        # GT 답 위치 분포 (참고용)
+        g_counter = Counter(g_positions)
+        print(f"\n  [GT 답의 위치 분포]")
+        print(f"  {'Position':<10} {'Count':<10} {'Ratio':<10}")
+        for i, pl in enumerate(position_labels):
+            cnt = g_counter.get(i, 0)
+            ratio = cnt / total * 100 if total > 0 else 0
+            print(f"  {pl:<10} {cnt:<10} {ratio:.1f}%")
+        g_std = np.std([g_counter.get(i, 0) / total * 100 for i in range(4)])
+        print(f"  Std: {g_std:.1f}%p")
+
+        # Heuristic == GT 비율
+        h_is_gt_total = heuristic_is_gt_far + heuristic_is_gt_close if label == 'FAR+CLOSE' else h_is_gt
+        print(f"\n  Heuristic == GT: {h_is_gt}/{total} ({h_is_gt/total*100:.1f}%)")
+        print(f"  → 이 비율이 Consistent 샘플 비율과 유사해야 함")
+
+    # ===== Heuristic 답 위치 vs GT 답 위치 교차 분석 =====
+    print(f"\n{'='*70}")
+    print("Heuristic 답 위치 vs GT 답 위치 교차 분석")
+    print(f"{'='*70}")
+
+    for label, h_positions, g_positions, total in [
+        ('FAR', heuristic_pos_far, gt_pos_far, total_far),
+        ('CLOSE', heuristic_pos_close, gt_pos_close, total_close),
+    ]:
+        print(f"\n  {label}: Heuristic 위치별 GT 위치 분포")
+        print(f"  (행: Heuristic 위치, 열: GT 위치)")
+        print(f"\n  {'Heur\\GT':<10}", end='')
+        for pl in position_labels:
+            print(f"{pl:<10}", end='')
+        print(f"{'Total':<10}")
+        print(f"  {'─'*50}")
+
+        cross = defaultdict(lambda: defaultdict(int))
+        for h, g in zip(h_positions, g_positions):
+            cross[h][g] += 1
+
+        for hi, hpl in enumerate(position_labels):
+            row_total = sum(cross[hi].values())
+            if row_total == 0:
+                continue
+            print(f"  {hpl:<10}", end='')
+            for gi in range(4):
+                cnt = cross[hi][gi]
+                pct = cnt / row_total * 100 if row_total > 0 else 0
+                print(f"{cnt}({pct:.0f}%){'':<2}", end='')
+            print(f"{row_total}")
+
+    # ===== 핵심 요약 =====
+    print(f"\n{'='*70}")
+    print("핵심 요약")
+    print(f"{'='*70}")
+
+    far_h_counter = Counter(heuristic_pos_far)
+    close_h_counter = Counter(heuristic_pos_close)
+
+    far_max_pos = max(range(4), key=lambda i: far_h_counter.get(i, 0))
+    far_max_pct = far_h_counter.get(far_max_pos, 0) / total_far * 100
+    close_max_pos = max(range(4), key=lambda i: close_h_counter.get(i, 0))
+    close_max_pct = close_h_counter.get(close_max_pos, 0) / total_close * 100
+
+    print(f"\n  FAR  heuristic 답 최다 위치: {position_labels[far_max_pos]} ({far_max_pct:.1f}%)")
+    print(f"  CLOSE heuristic 답 최다 위치: {position_labels[close_max_pos]} ({close_max_pct:.1f}%)")
+
+    far_d_pct = far_h_counter.get(3, 0) / total_far * 100
+    close_d_pct = close_h_counter.get(3, 0) / total_close * 100
+    print(f"\n  FAR  heuristic 답이 D 위치: {far_h_counter.get(3, 0)}/{total_far} ({far_d_pct:.1f}%)")
+    print(f"  CLOSE heuristic 답이 D 위치: {close_h_counter.get(3, 0)}/{total_close} ({close_d_pct:.1f}%)")
+
+    if far_d_pct > 30:
+        print(f"\n  ⚠ FAR에서 heuristic 답이 D에 {far_d_pct:.1f}% 편중!")
+        print(f"    → D bias 모델이 FAR에서 heuristic에 따라 D를 선택하는 경향 설명 가능")
+    else:
+        print(f"\n  FAR heuristic 답의 D 위치 비율이 균등({far_d_pct:.1f}%)이므로,")
+        print(f"  D bias가 FAR에서 더 심한 것은 선택지 배치 때문이 아닌 다른 요인일 가능성")
+
+    if args.output:
+        sys.stdout = tee.close()
+        print(f"Results saved to {args.output}")
+
+
+if __name__ == '__main__':
+    main()

answer_bias_results.txt

@@ -0,0 +1,1607 @@
+
+================================================================================
+Model: molmo-7B-O-0924
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     68.6%
+    B     912      25.1%     66.9%
+    C     852      23.4%     56.7%
+    D     918      25.2%     50.2%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     1102     30.3%     59.6%
+    B     1022     28.1%     59.7%
+    C     798      21.9%     60.5%
+    D     718      19.7%     64.2%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 4.32%p (uniform=0)
+    Overall Accuracy: 60.7%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     63.0%
+    B     290      24.0%     59.7%
+    C     289      24.0%     54.7%
+    D     308      25.5%     58.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     329      27.3%     61.1%
+    B     286      23.7%     60.5%
+    C     274      22.7%     57.7%
+    D     317      26.3%     56.5%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 1.85%p (uniform=0)
+    Overall Accuracy: 59.0%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     59.1%
+    B     156      26.3%     62.2%
+    C     130      21.9%     63.1%
+    D     149      25.1%     62.4%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     141      23.7%     66.7%
+    B     146      24.6%     66.4%
+    C     141      23.7%     58.2%
+    D     166      27.9%     56.0%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 1.74%p (uniform=0)
+    Overall Accuracy: 61.6%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     66.9%
+    B     134      21.9%     56.7%
+    C     159      26.0%     47.8%
+    D     159      26.0%     54.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     188      30.7%     56.9%
+    B     140      22.9%     54.3%
+    C     133      21.7%     57.1%
+    D     151      24.7%     57.0%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 3.46%p (uniform=0)
+    Overall Accuracy: 56.4%
+
+================================================================================
+Model: molmo-7B-O-0924-data_scale_exp_80k
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     58.7%
+    B     912      25.1%     51.6%
+    C     852      23.4%     60.4%
+    D     918      25.2%     41.2%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     1021     28.0%     55.0%
+    B     876      24.1%     53.8%
+    C     1035     28.4%     49.8%
+    D     708      19.5%     53.4%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 3.63%p (uniform=0)
+    Overall Accuracy: 52.9%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     58.3%
+    B     290      24.0%     53.4%
+    C     289      24.0%     56.4%
+    D     308      25.5%     52.6%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     315      26.1%     59.0%
+    B     287      23.8%     54.0%
+    C     285      23.6%     57.2%
+    D     319      26.5%     50.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 1.29%p (uniform=0)
+    Overall Accuracy: 55.2%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     49.7%
+    B     156      26.3%     55.8%
+    C     130      21.9%     53.1%
+    D     149      25.1%     49.7%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     141      23.7%     56.0%
+    B     163      27.4%     53.4%
+    C     128      21.5%     53.9%
+    D     162      27.3%     45.7%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 2.48%p (uniform=0)
+    Overall Accuracy: 52.0%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     66.9%
+    B     134      21.9%     50.7%
+    C     159      26.0%     59.1%
+    D     159      26.0%     55.3%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     174      28.4%     61.5%
+    B     124      20.3%     54.8%
+    C     157      25.7%     59.9%
+    D     157      25.7%     56.1%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 2.96%p (uniform=0)
+    Overall Accuracy: 58.3%
+
+================================================================================
+Model: molmo-7B-O-0924-data_scale_exp_400k
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     60.9%
+    B     912      25.1%     63.4%
+    C     852      23.4%     68.5%
+    D     918      25.2%     67.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     862      23.7%     67.6%
+    B     860      23.6%     67.2%
+    C     968      26.6%     60.3%
+    D     950      26.1%     64.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 1.36%p (uniform=0)
+    Overall Accuracy: 64.9%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     52.4%
+    B     290      24.0%     56.6%
+    C     289      24.0%     61.9%
+    D     308      25.5%     55.8%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     290      24.0%     57.6%
+    B     271      22.5%     60.5%
+    C     344      28.5%     52.0%
+    D     301      25.0%     57.1%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 2.22%p (uniform=0)
+    Overall Accuracy: 56.6%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     53.5%
+    B     156      26.3%     53.8%
+    C     130      21.9%     66.2%
+    D     149      25.1%     63.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     138      23.2%     61.6%
+    B     129      21.7%     65.1%
+    C     171      28.8%     50.3%
+    D     156      26.3%     60.3%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 2.73%p (uniform=0)
+    Overall Accuracy: 58.8%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     51.2%
+    B     134      21.9%     59.7%
+    C     159      26.0%     58.5%
+    D     159      26.0%     49.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     152      24.8%     53.9%
+    B     142      23.2%     56.3%
+    C     173      28.3%     53.8%
+    D     145      23.7%     53.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 1.98%p (uniform=0)
+    Overall Accuracy: 54.4%
+
+================================================================================
+Model: molmo-7B-O-0924-data_scale_exp_800k
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     59.3%
+    B     912      25.1%     63.8%
+    C     852      23.4%     75.2%
+    D     918      25.2%     78.9%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     723      19.9%     78.6%
+    B     769      21.1%     75.7%
+    C     984      27.0%     65.1%
+    D     1164     32.0%     62.2%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 4.85%p (uniform=0)
+    Overall Accuracy: 69.1%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     51.1%
+    B     290      24.0%     45.5%
+    C     289      24.0%     61.9%
+    D     308      25.5%     80.5%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     227      18.8%     71.8%
+    B     185      15.3%     71.4%
+    C     302      25.0%     59.3%
+    D     492      40.8%     50.4%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 9.76%p (uniform=0)
+    Overall Accuracy: 59.9%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     52.8%
+    B     156      26.3%     41.7%
+    C     130      21.9%     63.1%
+    D     149      25.1%     86.6%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     114      19.2%     73.7%
+    B     81       13.6%     80.2%
+    C     135      22.7%     60.7%
+    D     264      44.4%     48.9%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 11.68%p (uniform=0)
+    Overall Accuracy: 60.6%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     49.4%
+    B     134      21.9%     50.0%
+    C     159      26.0%     61.0%
+    D     159      26.0%     74.8%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     113      18.5%     69.9%
+    B     104      17.0%     64.4%
+    C     167      27.3%     58.1%
+    D     228      37.3%     52.2%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 8.10%p (uniform=0)
+    Overall Accuracy: 59.2%
+
+================================================================================
+Model: molmo-7B-O-0924-data_scale_exp_2m
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     70.9%
+    B     912      25.1%     72.4%
+    C     852      23.4%     72.1%
+    D     918      25.2%     81.8%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     860      23.6%     79.0%
+    B     824      22.6%     80.1%
+    C     807      22.2%     76.1%
+    D     1149     31.6%     65.4%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 3.83%p (uniform=0)
+    Overall Accuracy: 74.3%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     56.4%
+    B     290      24.0%     54.5%
+    C     289      24.0%     59.5%
+    D     308      25.5%     72.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     276      22.9%     65.2%
+    B     227      18.8%     69.6%
+    C     276      22.9%     62.3%
+    D     427      35.4%     52.0%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 6.23%p (uniform=0)
+    Overall Accuracy: 60.7%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     49.7%
+    B     156      26.3%     53.8%
+    C     130      21.9%     55.4%
+    D     149      25.1%     77.2%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     127      21.4%     62.2%
+    B     119      20.0%     70.6%
+    C     116      19.5%     62.1%
+    D     232      39.1%     49.6%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 8.14%p (uniform=0)
+    Overall Accuracy: 58.9%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     63.1%
+    B     134      21.9%     55.2%
+    C     159      26.0%     62.9%
+    D     159      26.0%     67.3%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     149      24.3%     67.8%
+    B     108      17.6%     68.5%
+    C     160      26.1%     62.5%
+    D     195      31.9%     54.9%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 5.07%p (uniform=0)
+    Overall Accuracy: 62.4%
+
+================================================================================
+Model: NVILA-Lite-2B
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     19.7%
+    B     912      25.1%     16.9%
+    C     852      23.4%     14.6%
+    D     918      25.2%     16.6%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     352      32.0%     53.7%
+    B     290      26.3%     53.1%
+    C     210      19.1%     59.0%
+    D     249      22.6%     61.0%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 4.77%p (uniform=0)
+    Overall Accuracy: 17.0%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     17.6%
+    B     290      24.0%     13.8%
+    C     289      24.0%     11.4%
+    D     308      25.5%     14.3%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     120      33.0%     46.7%
+    B     91       25.0%     44.0%
+    C     71       19.5%     46.5%
+    D     82       22.5%     53.7%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 4.99%p (uniform=0)
+    Overall Accuracy: 14.3%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     18.2%
+    B     156      26.3%     12.8%
+    C     130      21.9%     11.5%
+    D     149      25.1%     11.4%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     65       36.5%     44.6%
+    B     37       20.8%     54.1%
+    C     32       18.0%     46.9%
+    D     44       24.7%     38.6%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 7.07%p (uniform=0)
+    Overall Accuracy: 13.6%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     16.9%
+    B     134      21.9%     14.9%
+    C     159      26.0%     11.3%
+    D     159      26.0%     17.0%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     55       29.6%     49.1%
+    B     54       29.0%     37.0%
+    C     39       21.0%     46.2%
+    D     38       20.4%     71.1%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 4.31%p (uniform=0)
+    Overall Accuracy: 15.0%
+
+================================================================================
+Model: NVILA-Lite-2B-data-scale-exp-80k
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     63.9%
+    B     912      25.1%     69.2%
+    C     852      23.4%     61.7%
+    D     918      25.2%     65.5%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     930      25.5%     65.8%
+    B     1010     27.7%     62.5%
+    C     786      21.6%     66.9%
+    D     914      25.1%     65.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 2.21%p (uniform=0)
+    Overall Accuracy: 65.1%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     58.3%
+    B     290      24.0%     52.8%
+    C     289      24.0%     43.9%
+    D     308      25.5%     46.4%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     363      30.1%     51.2%
+    B     337      27.9%     45.4%
+    C     234      19.4%     54.3%
+    D     272      22.6%     52.6%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 4.24%p (uniform=0)
+    Overall Accuracy: 50.5%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     64.8%
+    B     156      26.3%     58.3%
+    C     130      21.9%     46.2%
+    D     149      25.1%     51.0%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     194      32.7%     53.1%
+    B     171      28.8%     53.2%
+    C     108      18.2%     55.6%
+    D     121      20.4%     62.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 5.94%p (uniform=0)
+    Overall Accuracy: 55.6%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     51.9%
+    B     134      21.9%     46.3%
+    C     159      26.0%     42.1%
+    D     159      26.0%     42.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     169      27.6%     49.1%
+    B     166      27.1%     37.3%
+    C     126      20.6%     53.2%
+    D     151      24.7%     44.4%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 2.78%p (uniform=0)
+    Overall Accuracy: 45.6%
+
+================================================================================
+Model: NVILA-Lite-2B-data-scale-exp-400k
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     58.9%
+    B     912      25.1%     60.0%
+    C     852      23.4%     65.3%
+    D     918      25.2%     64.5%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     889      24.4%     63.4%
+    B     849      23.3%     64.4%
+    C     905      24.9%     61.4%
+    D     997      27.4%     59.4%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 1.49%p (uniform=0)
+    Overall Accuracy: 62.1%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     58.9%
+    B     290      24.0%     53.4%
+    C     289      24.0%     59.5%
+    D     308      25.5%     52.9%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     341      28.3%     55.1%
+    B     267      22.1%     58.1%
+    C     303      25.1%     56.8%
+    D     295      24.5%     55.3%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 2.19%p (uniform=0)
+    Overall Accuracy: 56.2%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     62.3%
+    B     156      26.3%     56.4%
+    C     130      21.9%     63.8%
+    D     149      25.1%     53.0%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     177      29.8%     55.9%
+    B     138      23.2%     63.8%
+    C     146      24.6%     56.8%
+    D     133      22.4%     59.4%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 2.88%p (uniform=0)
+    Overall Accuracy: 58.8%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     55.6%
+    B     134      21.9%     50.0%
+    C     159      26.0%     56.0%
+    D     159      26.0%     52.8%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     164      26.8%     54.3%
+    B     129      21.1%     51.9%
+    C     157      25.7%     56.7%
+    D     162      26.5%     51.9%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 2.30%p (uniform=0)
+    Overall Accuracy: 53.8%
+
+================================================================================
+Model: NVILA-Lite-2B-data-scale-exp-800k
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     68.7%
+    B     912      25.1%     65.7%
+    C     852      23.4%     71.6%
+    D     918      25.2%     73.0%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     931      25.6%     70.7%
+    B     794      21.8%     75.4%
+    C     893      24.5%     68.3%
+    D     1022     28.1%     65.6%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 2.25%p (uniform=0)
+    Overall Accuracy: 69.7%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     65.5%
+    B     290      24.0%     50.7%
+    C     289      24.0%     58.1%
+    D     308      25.5%     57.8%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     381      31.6%     54.9%
+    B     236      19.6%     62.3%
+    C     283      23.5%     59.4%
+    D     306      25.4%     58.2%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 4.34%p (uniform=0)
+    Overall Accuracy: 58.2%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     71.7%
+    B     156      26.3%     55.1%
+    C     130      21.9%     62.3%
+    D     149      25.1%     61.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     199      33.5%     57.3%
+    B     121      20.4%     71.1%
+    C     127      21.4%     63.8%
+    D     147      24.7%     61.9%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 5.17%p (uniform=0)
+    Overall Accuracy: 62.6%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     59.4%
+    B     134      21.9%     45.5%
+    C     159      26.0%     54.7%
+    D     159      26.0%     54.7%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     182      29.7%     52.2%
+    B     115      18.8%     53.0%
+    C     156      25.5%     55.8%
+    D     159      26.0%     54.7%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 3.94%p (uniform=0)
+    Overall Accuracy: 53.9%
+
+================================================================================
+Model: NVILA-Lite-2B-data-scale-exp-2m
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     72.4%
+    B     912      25.1%     65.1%
+    C     852      23.4%     70.2%
+    D     918      25.2%     69.6%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     1008     27.7%     68.8%
+    B     784      21.5%     75.8%
+    C     892      24.5%     67.0%
+    D     956      26.3%     66.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 2.30%p (uniform=0)
+    Overall Accuracy: 69.4%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     64.9%
+    B     290      24.0%     50.7%
+    C     289      24.0%     58.5%
+    D     308      25.5%     52.9%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     379      31.4%     54.6%
+    B     239      19.8%     61.5%
+    C     305      25.3%     55.4%
+    D     283      23.5%     57.6%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 4.20%p (uniform=0)
+    Overall Accuracy: 56.9%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     69.8%
+    B     156      26.3%     55.8%
+    C     130      21.9%     64.6%
+    D     149      25.1%     55.0%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     194      32.7%     57.2%
+    B     126      21.2%     69.0%
+    C     144      24.2%     58.3%
+    D     130      21.9%     63.1%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 4.56%p (uniform=0)
+    Overall Accuracy: 61.3%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     60.0%
+    B     134      21.9%     44.8%
+    C     159      26.0%     53.5%
+    D     159      26.0%     50.9%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     185      30.2%     51.9%
+    B     113      18.5%     53.1%
+    C     161      26.3%     52.8%
+    D     153      25.0%     52.9%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 4.24%p (uniform=0)
+    Overall Accuracy: 52.6%
+
+================================================================================
+Model: RoboRefer-2B-SFT
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     94.1%
+    B     912      25.1%     93.9%
+    C     852      23.4%     92.1%
+    D     918      25.2%     88.0%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     991      27.3%     90.8%
+    B     937      25.8%     91.2%
+    C     851      23.4%     91.9%
+    D     851      23.4%     94.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 1.64%p (uniform=0)
+    Overall Accuracy: 92.0%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     88.1%
+    B     290      24.0%     84.8%
+    C     289      24.0%     81.7%
+    D     308      25.5%     76.0%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     352      29.4%     79.5%
+    B     295      24.7%     83.1%
+    C     280      23.4%     83.2%
+    D     269      22.5%     86.6%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 2.67%p (uniform=0)
+    Overall Accuracy: 82.7%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     90.6%
+    B     156      26.3%     85.3%
+    C     130      21.9%     76.2%
+    D     149      25.1%     72.5%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     186      31.4%     76.9%
+    B     161      27.2%     82.0%
+    C     119      20.1%     82.4%
+    D     126      21.3%     86.5%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 4.58%p (uniform=0)
+    Overall Accuracy: 81.5%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     85.6%
+    B     134      21.9%     84.3%
+    C     159      26.0%     86.2%
+    D     159      26.0%     79.2%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     166      27.5%     82.5%
+    B     134      22.2%     84.3%
+    C     161      26.7%     83.9%
+    D     143      23.7%     86.7%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 2.16%p (uniform=0)
+    Overall Accuracy: 83.8%
+
+================================================================================
+Model: Qwen2.5-VL-3B-Instruct
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     51.9%
+    B     912      25.1%     61.5%
+    C     852      23.4%     64.0%
+    D     918      25.2%     72.3%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     682      18.7%     72.9%
+    B     863      23.7%     65.0%
+    C     938      25.8%     58.1%
+    D     1157     31.8%     57.4%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 4.68%p (uniform=0)
+    Overall Accuracy: 62.3%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     42.3%
+    B     290      24.0%     47.2%
+    C     289      24.0%     49.8%
+    D     308      25.5%     62.7%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     226      18.7%     59.7%
+    B     266      22.1%     51.5%
+    C     301      25.0%     47.8%
+    D     413      34.2%     46.7%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 5.77%p (uniform=0)
+    Overall Accuracy: 50.5%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     42.8%
+    B     156      26.3%     48.1%
+    C     130      21.9%     48.5%
+    D     149      25.1%     66.4%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     114      19.2%     59.6%
+    B     135      22.7%     55.6%
+    C     134      22.6%     47.0%
+    D     211      35.5%     46.9%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 6.24%p (uniform=0)
+    Overall Accuracy: 51.3%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     41.9%
+    B     134      21.9%     46.3%
+    C     159      26.0%     50.9%
+    D     159      26.0%     59.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     112      18.3%     59.8%
+    B     131      21.4%     47.3%
+    C     167      27.3%     48.5%
+    D     202      33.0%     46.5%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 5.64%p (uniform=0)
+    Overall Accuracy: 49.7%
+
+================================================================================
+Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_80k
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     47.9%
+    B     912      25.1%     57.0%
+    C     852      23.4%     61.6%
+    D     918      25.2%     63.5%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     704      19.3%     65.2%
+    B     883      24.3%     58.9%
+    C     974      26.8%     53.9%
+    D     1079     29.6%     54.0%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 3.78%p (uniform=0)
+    Overall Accuracy: 57.3%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     41.7%
+    B     290      24.0%     45.5%
+    C     289      24.0%     45.3%
+    D     308      25.5%     54.5%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     238      19.7%     55.9%
+    B     270      22.4%     48.9%
+    C     296      24.5%     44.3%
+    D     402      33.3%     41.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 5.10%p (uniform=0)
+    Overall Accuracy: 46.8%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     44.0%
+    B     156      26.3%     46.8%
+    C     130      21.9%     46.2%
+    D     149      25.1%     61.7%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     123      20.7%     56.9%
+    B     131      22.1%     55.7%
+    C     139      23.4%     43.2%
+    D     201      33.8%     45.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 5.19%p (uniform=0)
+    Overall Accuracy: 49.7%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     39.4%
+    B     134      21.9%     44.0%
+    C     159      26.0%     44.7%
+    D     159      26.0%     47.8%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     115      18.8%     54.8%
+    B     139      22.7%     42.4%
+    C     157      25.7%     45.2%
+    D     201      32.8%     37.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 5.14%p (uniform=0)
+    Overall Accuracy: 44.0%
+
+================================================================================
+Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_400k
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     50.1%
+    B     912      25.1%     56.6%
+    C     852      23.4%     62.3%
+    D     918      25.2%     66.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     736      20.2%     65.2%
+    B     825      22.7%     62.5%
+    C     966      26.5%     55.0%
+    D     1113     30.6%     54.5%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 3.93%p (uniform=0)
+    Overall Accuracy: 58.6%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     43.6%
+    B     290      24.0%     43.4%
+    C     289      24.0%     45.3%
+    D     308      25.5%     60.4%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     243      20.1%     57.2%
+    B     238      19.7%     52.9%
+    C     299      24.8%     43.8%
+    D     426      35.3%     43.7%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 6.28%p (uniform=0)
+    Overall Accuracy: 48.3%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     46.5%
+    B     156      26.3%     44.9%
+    C     130      21.9%     48.5%
+    D     149      25.1%     65.8%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     126      21.2%     58.7%
+    B     118      19.9%     59.3%
+    C     143      24.1%     44.1%
+    D     207      34.8%     47.3%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 5.89%p (uniform=0)
+    Overall Accuracy: 51.3%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     40.6%
+    B     134      21.9%     41.8%
+    C     159      26.0%     42.8%
+    D     159      26.0%     55.3%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     117      19.1%     55.6%
+    B     120      19.6%     46.7%
+    C     156      25.5%     43.6%
+    D     219      35.8%     40.2%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 6.71%p (uniform=0)
+    Overall Accuracy: 45.3%
+
+================================================================================
+Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_800k
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     55.7%
+    B     912      25.1%     57.9%
+    C     852      23.4%     63.5%
+    D     918      25.2%     66.7%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     814      22.4%     65.6%
+    B     810      22.3%     65.2%
+    C     937      25.7%     57.7%
+    D     1079     29.6%     56.7%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 3.03%p (uniform=0)
+    Overall Accuracy: 60.9%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     51.4%
+    B     290      24.0%     45.9%
+    C     289      24.0%     45.7%
+    D     308      25.5%     58.8%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     294      24.4%     55.8%
+    B     235      19.5%     56.6%
+    C     283      23.5%     46.6%
+    D     394      32.7%     45.9%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 4.80%p (uniform=0)
+    Overall Accuracy: 50.6%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     56.6%
+    B     156      26.3%     46.8%
+    C     130      21.9%     50.0%
+    D     149      25.1%     65.1%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     156      26.3%     57.7%
+    B     109      18.4%     67.0%
+    C     139      23.4%     46.8%
+    D     190      32.0%     51.1%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 4.93%p (uniform=0)
+    Overall Accuracy: 54.7%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     46.2%
+    B     134      21.9%     44.8%
+    C     159      26.0%     42.1%
+    D     159      26.0%     52.8%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     138      22.5%     53.6%
+    B     126      20.6%     47.6%
+    C     144      23.5%     46.5%
+    D     204      33.3%     41.2%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 4.93%p (uniform=0)
+    Overall Accuracy: 46.6%
+
+================================================================================
+Model: Qwen2.5-VL-3B-Instruct-data_scale_exp_2m
+================================================================================
+
+--- ALL (n=3640) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     958      26.3%     63.3%
+    B     912      25.1%     63.7%
+    C     852      23.4%     69.8%
+    D     918      25.2%     66.6%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     890      24.5%     68.1%
+    B     831      22.8%     69.9%
+    C     961      26.4%     61.9%
+    D     958      26.3%     63.8%
+
+  Bias Indicators:
+    GT Distribution Std: 1.04%p (uniform=0)
+    Pred Distribution Std: 1.48%p (uniform=0)
+    Overall Accuracy: 65.7%
+
+--- FAR+CLOSE (n=1206) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     319      26.5%     57.4%
+    B     290      24.0%     50.0%
+    C     289      24.0%     54.7%
+    D     308      25.5%     55.8%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     322      26.7%     56.8%
+    B     245      20.3%     59.2%
+    C     309      25.6%     51.1%
+    D     330      27.4%     52.1%
+
+  Bias Indicators:
+    GT Distribution Std: 1.05%p (uniform=0)
+    Pred Distribution Std: 2.78%p (uniform=0)
+    Overall Accuracy: 54.6%
+
+--- FAR (n=594) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     159      26.8%     61.0%
+    B     156      26.3%     50.0%
+    C     130      21.9%     53.1%
+    D     149      25.1%     59.7%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     165      27.8%     58.8%
+    B     127      21.4%     61.4%
+    C     146      24.6%     47.3%
+    D     156      26.3%     57.1%
+
+  Bias Indicators:
+    GT Distribution Std: 1.90%p (uniform=0)
+    Pred Distribution Std: 2.38%p (uniform=0)
+    Overall Accuracy: 56.1%
+
+--- CLOSE (n=612) ---
+
+  GT Answer Distribution:
+    Pos   Count    Pct      Acc when GT 
+    -----------------------------------
+    A     160      26.1%     53.8%
+    B     134      21.9%     50.0%
+    C     159      26.0%     56.0%
+    D     159      26.0%     52.2%
+
+  Model Prediction Distribution:
+    Pos   Count    Pct      Acc when Pred
+    -----------------------------------
+    A     157      25.7%     54.8%
+    B     118      19.3%     56.8%
+    C     163      26.6%     54.6%
+    D     174      28.4%     47.7%
+
+  Bias Indicators:
+    GT Distribution Std: 1.79%p (uniform=0)
+    Pred Distribution Std: 3.45%p (uniform=0)
+    Overall Accuracy: 53.1%
+
+====================================================================================================
+MODEL BIAS COMPARISON SUMMARY
+====================================================================================================
+
+Model                                         Subset       GT Std     Pred Std   Pred Max     Acc     
+-------------------------------------------------------------------------------------------------
+molmo-7B-O-0924                               ALL          1.0%p     4.3%p     A(30.3%)   60.7%
+                                              FAR+CLOSE    1.0%p     1.9%p     A(27.3%)   59.0%
+                                              FAR          1.9%p     1.7%p     D(27.9%)   61.6%
+                                              CLOSE        1.8%p     3.5%p     A(30.7%)   56.4%
+molmo-7B-O-0924-data_scale_exp_80k            ALL          1.0%p     3.6%p     C(28.4%)   52.9%
+                                              FAR+CLOSE    1.0%p     1.3%p     D(26.5%)   55.2%
+                                              FAR          1.9%p     2.5%p     B(27.4%)   52.0%
+                                              CLOSE        1.8%p     3.0%p     A(28.4%)   58.3%
+molmo-7B-O-0924-data_scale_exp_400k           ALL          1.0%p     1.4%p     C(26.6%)   64.9%
+                                              FAR+CLOSE    1.0%p     2.2%p     C(28.5%)   56.6%
+                                              FAR          1.9%p     2.7%p     C(28.8%)   58.8%
+                                              CLOSE        1.8%p     2.0%p     C(28.3%)   54.4%
+molmo-7B-O-0924-data_scale_exp_800k           ALL          1.0%p     4.9%p     D(32.0%)   69.1%
+                                              FAR+CLOSE    1.0%p     9.8%p     D(40.8%)   59.9%
+                                              FAR          1.9%p     11.7%p    D(44.4%)   60.6%
+                                              CLOSE        1.8%p     8.1%p     D(37.3%)   59.2%
+molmo-7B-O-0924-data_scale_exp_2m             ALL          1.0%p     3.8%p     D(31.6%)   74.3%
+                                              FAR+CLOSE    1.0%p     6.2%p     D(35.4%)   60.7%
+                                              FAR          1.9%p     8.1%p     D(39.1%)   58.9%
+                                              CLOSE        1.8%p     5.1%p     D(31.9%)   62.4%
+NVILA-Lite-2B                                 ALL          1.0%p     4.8%p     A(32.0%)   17.0%
+                                              FAR+CLOSE    1.0%p     5.0%p     A(33.0%)   14.3%
+                                              FAR          1.9%p     7.1%p     A(36.5%)   13.6%
+                                              CLOSE        1.8%p     4.3%p     A(29.6%)   15.0%
+NVILA-Lite-2B-data-scale-exp-80k              ALL          1.0%p     2.2%p     B(27.7%)   65.1%
+                                              FAR+CLOSE    1.0%p     4.2%p     A(30.1%)   50.5%
+                                              FAR          1.9%p     5.9%p     A(32.7%)   55.6%
+                                              CLOSE        1.8%p     2.8%p     A(27.6%)   45.6%
+NVILA-Lite-2B-data-scale-exp-400k             ALL          1.0%p     1.5%p     D(27.4%)   62.1%
+                                              FAR+CLOSE    1.0%p     2.2%p     A(28.3%)   56.2%
+                                              FAR          1.9%p     2.9%p     A(29.8%)   58.8%
+                                              CLOSE        1.8%p     2.3%p     A(26.8%)   53.8%
+NVILA-Lite-2B-data-scale-exp-800k             ALL          1.0%p     2.2%p     D(28.1%)   69.7%
+                                              FAR+CLOSE    1.0%p     4.3%p     A(31.6%)   58.2%
+                                              FAR          1.9%p     5.2%p     A(33.5%)   62.6%
+                                              CLOSE        1.8%p     3.9%p     A(29.7%)   53.9%
+NVILA-Lite-2B-data-scale-exp-2m               ALL          1.0%p     2.3%p     A(27.7%)   69.4%
+                                              FAR+CLOSE    1.0%p     4.2%p     A(31.4%)   56.9%
+                                              FAR          1.9%p     4.6%p     A(32.7%)   61.3%
+                                              CLOSE        1.8%p     4.2%p     A(30.2%)   52.6%
+RoboRefer-2B-SFT                              ALL          1.0%p     1.6%p     A(27.3%)   92.0%
+                                              FAR+CLOSE    1.0%p     2.7%p     A(29.4%)   82.7%
+                                              FAR          1.9%p     4.6%p     A(31.4%)   81.5%
+                                              CLOSE        1.8%p     2.2%p     A(27.5%)   83.8%
+Qwen2.5-VL-3B-Instruct                        ALL          1.0%p     4.7%p     D(31.8%)   62.3%
+                                              FAR+CLOSE    1.0%p     5.8%p     D(34.2%)   50.5%
+                                              FAR          1.9%p     6.2%p     D(35.5%)   51.3%
+                                              CLOSE        1.8%p     5.6%p     D(33.0%)   49.7%
+Qwen2.5-VL-3B-Instruct-data_scale_exp_80k     ALL          1.0%p     3.8%p     D(29.6%)   57.3%
+                                              FAR+CLOSE    1.0%p     5.1%p     D(33.3%)   46.8%
+                                              FAR          1.9%p     5.2%p     D(33.8%)   49.7%
+                                              CLOSE        1.8%p     5.1%p     D(32.8%)   44.0%
+Qwen2.5-VL-3B-Instruct-data_scale_exp_400k    ALL          1.0%p     3.9%p     D(30.6%)   58.6%
+                                              FAR+CLOSE    1.0%p     6.3%p     D(35.3%)   48.3%
+                                              FAR          1.9%p     5.9%p     D(34.8%)   51.3%
+                                              CLOSE        1.8%p     6.7%p     D(35.8%)   45.3%
+Qwen2.5-VL-3B-Instruct-data_scale_exp_800k    ALL          1.0%p     3.0%p     D(29.6%)   60.9%
+                                              FAR+CLOSE    1.0%p     4.8%p     D(32.7%)   50.6%
+                                              FAR          1.9%p     4.9%p     D(32.0%)   54.7%
+                                              CLOSE        1.8%p     4.9%p     D(33.3%)   46.6%
+Qwen2.5-VL-3B-Instruct-data_scale_exp_2m      ALL          1.0%p     1.5%p     C(26.4%)   65.7%
+                                              FAR+CLOSE    1.0%p     2.8%p     D(27.4%)   54.6%
+                                              FAR          1.9%p     2.4%p     A(27.8%)   56.1%
+                                              CLOSE        1.8%p     3.4%p     D(28.4%)   53.1%

correct_filter/results/nvila/correct_only/csv/similarity_2m_L0.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999998,0.99979234,0.9990382,0.99845904,0.99873054,0.99917614
+right,0.99979234,1.0000002,0.99917865,0.9987631,0.9989448,0.9993073
+above,0.9990382,0.99917865,0.9999999,0.9996766,0.99956864,0.9995194
+under,0.99845904,0.9987631,0.9996766,0.99999976,0.9995517,0.9991601
+far,0.99873054,0.9989448,0.99956864,0.9995517,1.0000004,0.9997673
+close,0.99917614,0.9993073,0.9995194,0.9991601,0.9997673,1.0000001

correct_filter/results/nvila/correct_only/csv/similarity_2m_L10.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000002,0.99946415,0.99536633,0.9942555,0.9812977,0.98185515
+right,0.99946415,1.0000001,0.9956668,0.9950104,0.98220986,0.9825698
+above,0.99536633,0.9956668,1.0000004,0.9991444,0.98727834,0.9875144
+under,0.9942555,0.9950104,0.9991444,1.0000002,0.98638064,0.9860885
+far,0.9812977,0.98220986,0.98727834,0.98638064,1.0,0.9996808
+close,0.98185515,0.9825698,0.9875144,0.9860885,0.9996808,0.9999997

correct_filter/results/nvila/correct_only/csv/similarity_2m_L12.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.99999994,0.9992977,0.9910708,0.9898012,0.97599757,0.9766798
+right,0.9992977,0.9999998,0.9917161,0.9909911,0.97698414,0.9776015
+above,0.9910708,0.9917161,1.0,0.9988291,0.98402643,0.98436046
+under,0.9898012,0.9909911,0.9988291,0.9999997,0.98302853,0.9827963
+far,0.97599757,0.97698414,0.98402643,0.98302853,0.99999976,0.9993965
+close,0.9766798,0.9776015,0.98436046,0.9827963,0.9993965,0.99999934

correct_filter/results/nvila/correct_only/csv/similarity_2m_L15.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000006,0.9960951,0.9754069,0.9750978,0.9668617,0.9666722
+right,0.9960951,1.0000005,0.9735367,0.9757243,0.9667201,0.966353
+above,0.9754069,0.9735367,1.000001,0.99429655,0.9720154,0.9716815
+under,0.9750978,0.9757243,0.99429655,0.99999964,0.9723926,0.9707299
+far,0.9668617,0.9667201,0.9720154,0.9723926,0.9999996,0.9988003
+close,0.9666722,0.966353,0.9716815,0.9707299,0.9988003,1.0000001

correct_filter/results/nvila/correct_only/csv/similarity_2m_L20.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0,0.8997723,0.8560219,0.84100085,0.812119,0.81015867
+right,0.8997723,1.0,0.8408875,0.8583672,0.8323572,0.82982045
+above,0.8560219,0.8408875,0.99999976,0.83596146,0.8394874,0.81615543
+under,0.84100085,0.8583672,0.83596146,1.0000001,0.8413338,0.8540078
+far,0.812119,0.8323572,0.8394874,0.8413338,1.0000002,0.9614612
+close,0.81015867,0.82982045,0.81615543,0.8540078,0.9614612,1.0000004

correct_filter/results/nvila/correct_only/csv/similarity_2m_L21.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.99999964,0.9202951,0.8803115,0.8690176,0.844957,0.8430869
+right,0.9202951,0.9999999,0.86974925,0.8861736,0.8690333,0.8677544
+above,0.8803115,0.86974925,0.9999997,0.8800344,0.8820302,0.865402
+under,0.8690176,0.8861736,0.8800344,1.0,0.8819223,0.8901774
+far,0.844957,0.8690333,0.8820302,0.8819223,1.0000001,0.9756321
+close,0.8430869,0.8677544,0.865402,0.8901774,0.9756321,0.99999934

correct_filter/results/nvila/correct_only/csv/similarity_2m_L24.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999998,0.9441167,0.8852379,0.877577,0.8462348,0.84368265
+right,0.9441167,0.9999995,0.86733294,0.88120615,0.8449521,0.8402463
+above,0.8852379,0.86733294,1.0,0.9256213,0.8766693,0.86528826
+under,0.877577,0.88120615,0.9256213,0.99999964,0.8739595,0.8758325
+far,0.8462348,0.8449521,0.8766693,0.8739595,1.0,0.9843055
+close,0.84368265,0.8402463,0.86528826,0.8758325,0.9843055,0.9999998

correct_filter/results/nvila/correct_only/csv/similarity_2m_L4.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0,0.99861413,0.99788797,0.99436826,0.9940919,0.9947342
+right,0.99861413,1.0000002,0.99827963,0.9968713,0.9948081,0.9949278
+above,0.99788797,0.99827963,1.0000002,0.99800307,0.9956355,0.9955324
+under,0.99436826,0.9968713,0.99800307,1.0000001,0.9943704,0.99312276
+far,0.9940919,0.9948081,0.9956355,0.9943704,1.0,0.9994365
+close,0.9947342,0.9949278,0.9955324,0.99312276,0.9994365,1.0000002

correct_filter/results/nvila/correct_only/csv/similarity_400k_L0.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.99999964,0.99937326,0.99939746,0.9992237,0.99935555,0.99926114
+right,0.99937326,1.0,0.99907005,0.9993704,0.9991206,0.9992242
+above,0.99939746,0.99907005,1.0000002,0.999896,0.99969333,0.9996633
+under,0.9992237,0.9993704,0.999896,1.0000002,0.9996517,0.99970394
+far,0.99935555,0.9991206,0.99969333,0.9996517,0.9999995,0.99996144
+close,0.99926114,0.9992242,0.9996633,0.99970394,0.99996144,0.99999946

correct_filter/results/nvila/correct_only/csv/similarity_400k_L10.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.99999994,0.9989206,0.9917971,0.99177206,0.9653694,0.9650217
+right,0.9989206,0.99999994,0.9908831,0.991404,0.96409506,0.9640688
+above,0.9917971,0.9908831,1.0,0.9998317,0.98056906,0.98046494
+under,0.99177206,0.991404,0.9998317,1.0,0.9801248,0.9800962
+far,0.9653694,0.96409506,0.98056906,0.9801248,0.99999976,0.99993694
+close,0.9650217,0.9640688,0.98046494,0.9800962,0.99993694,1.0000001

correct_filter/results/nvila/correct_only/csv/similarity_400k_L12.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999998,0.9988699,0.98620236,0.9862741,0.961798,0.9616283
+right,0.9988699,0.99999994,0.9853998,0.98599064,0.9603799,0.9605291
+above,0.98620236,0.9853998,0.9999995,0.99963886,0.9798792,0.9798204
+under,0.9862741,0.98599064,0.99963886,0.9999998,0.97985244,0.9798428
+far,0.961798,0.9603799,0.9798792,0.97985244,0.9999997,0.9998978
+close,0.9616283,0.9605291,0.9798204,0.9798428,0.9998978,1.0000001

correct_filter/results/nvila/correct_only/csv/similarity_400k_L13.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000001,0.9983739,0.98168665,0.98217195,0.95686316,0.956294
+right,0.9983739,0.9999995,0.98163897,0.9828724,0.9566915,0.9564952
+above,0.98168665,0.98163897,0.99999917,0.9987243,0.975779,0.9747824
+under,0.98217195,0.9828724,0.9987243,0.99999976,0.9764465,0.9758391
+far,0.95686316,0.9566915,0.975779,0.9764465,1.0000002,0.99971694
+close,0.956294,0.9564952,0.9747824,0.9758391,0.99971694,1.0

correct_filter/results/nvila/correct_only/csv/similarity_400k_L14.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999998,0.9985205,0.98138416,0.9822928,0.9568843,0.9561093
+right,0.9985205,1.0000004,0.98126113,0.9829378,0.95713615,0.9566823
+above,0.98138416,0.98126113,0.9999999,0.9986111,0.9755904,0.97441
+under,0.9822928,0.9829378,0.9986111,1.0000001,0.9771458,0.9763574
+far,0.9568843,0.95713615,0.9755904,0.9771458,0.9999996,0.9997202
+close,0.9561093,0.9566823,0.97441,0.9763574,0.9997202,0.9999998

correct_filter/results/nvila/correct_only/csv/similarity_400k_L15.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.99999976,0.9961381,0.96402776,0.9712392,0.94819504,0.94753337
+right,0.9961381,0.99999946,0.96187425,0.9719884,0.94920367,0.9488889
+above,0.96402776,0.96187425,0.99999994,0.99324864,0.95932007,0.9581619
+under,0.9712392,0.9719884,0.99324864,1.0000002,0.9649157,0.9638857
+far,0.94819504,0.94920367,0.95932007,0.9649157,0.99999964,0.9996454
+close,0.94753337,0.9488889,0.9581619,0.9638857,0.9996454,0.9999997

correct_filter/results/nvila/correct_only/csv/similarity_400k_L20.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000002,0.94959533,0.9076828,0.8961903,0.8253381,0.8177571
+right,0.94959533,1.0000001,0.89672065,0.9084783,0.8551606,0.85423845
+above,0.9076828,0.89672065,0.9999996,0.9012549,0.8612656,0.83851355
+under,0.8961903,0.9084783,0.9012549,0.99999964,0.8549114,0.87088656
+far,0.8253381,0.8551606,0.8612656,0.8549114,0.9999998,0.97840255
+close,0.8177571,0.85423845,0.83851355,0.87088656,0.97840255,1.0

correct_filter/results/nvila/correct_only/csv/similarity_400k_L21.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.99999976,0.9635052,0.92306554,0.9076893,0.8688898,0.86037034
+right,0.9635052,1.0000005,0.9213424,0.9254695,0.90012884,0.8971458
+above,0.92306554,0.9213424,1.0000001,0.9205528,0.89912474,0.88151306
+under,0.9076893,0.9254695,0.9205528,1.0000001,0.88736314,0.8975957
+far,0.8688898,0.90012884,0.89912474,0.88736314,1.0000001,0.98656934
+close,0.86037034,0.8971458,0.88151306,0.8975957,0.98656934,1.0

correct_filter/results/nvila/correct_only/csv/similarity_400k_L22.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000002,0.96869767,0.920773,0.9017892,0.86359686,0.85843223
+right,0.96869767,0.9999998,0.91396976,0.9125042,0.8843721,0.88239276
+above,0.920773,0.91396976,0.9999998,0.93451786,0.87860537,0.8666265
+under,0.9017892,0.9125042,0.93451786,1.0000002,0.8649305,0.8732993
+far,0.86359686,0.8843721,0.87860537,0.8649305,0.9999998,0.98876476
+close,0.85843223,0.88239276,0.8666265,0.8732993,0.98876476,1.0000001

correct_filter/results/nvila/correct_only/csv/similarity_400k_L27.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000002,0.99905974,0.9938745,0.9942479,0.9916425,0.99195606
+right,0.99905974,1.0000002,0.99394035,0.9950758,0.9930149,0.9933292
+above,0.9938745,0.99394035,1.0000004,0.997444,0.9924405,0.9923437
+under,0.9942479,0.9950758,0.997444,0.99999964,0.99323577,0.9936469
+far,0.9916425,0.9930149,0.9924405,0.99323577,0.9999994,0.9997259
+close,0.99195606,0.9933292,0.9923437,0.9936469,0.9997259,1.0000001

correct_filter/results/nvila/correct_only/csv/similarity_400k_L5.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999999,0.9972796,0.9979343,0.9974654,0.9934282,0.9930154
+right,0.9972796,1.0,0.996982,0.99778414,0.9929778,0.9934241
+above,0.9979343,0.996982,0.9999998,0.99976426,0.994526,0.9944363
+under,0.9974654,0.99778414,0.99976426,1.0000001,0.9946017,0.9946864
+far,0.9934282,0.9929778,0.994526,0.9946017,1.0000001,0.99988496
+close,0.9930154,0.9934241,0.9944363,0.9946864,0.99988496,0.9999999

correct_filter/results/nvila/correct_only/csv/similarity_400k_L6.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.99999994,0.99787366,0.99845725,0.99814403,0.9947071,0.9944121
+right,0.99787366,1.0000002,0.9980315,0.99859196,0.99486274,0.9951682
+above,0.99845725,0.9980315,0.99999964,0.99983656,0.9958496,0.99578476
+under,0.99814403,0.99859196,0.99983656,0.99999946,0.99596035,0.99600405
+far,0.9947071,0.99486274,0.9958496,0.99596035,0.9999999,0.9999253
+close,0.9944121,0.9951682,0.99578476,0.99600405,0.9999253,1.0000005

correct_filter/results/nvila/correct_only/csv/similarity_400k_L7.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999998,0.9982521,0.99867207,0.998369,0.9958096,0.9954878
+right,0.9982521,0.99999917,0.99839646,0.9988315,0.99623215,0.9964118
+above,0.99867207,0.99839646,1.0,0.99986494,0.9967103,0.9965906
+under,0.998369,0.9988315,0.99986494,0.9999998,0.99682885,0.99680567
+far,0.9958096,0.99623215,0.9967103,0.99682885,0.9999999,0.9999416
+close,0.9954878,0.9964118,0.9965906,0.99680567,0.9999416,1.0000002

correct_filter/results/nvila/correct_only/csv/similarity_800k_L10.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999998,0.9997335,0.9946324,0.9929318,0.98346335,0.9839893
+right,0.9997335,0.99999976,0.9943074,0.99233896,0.98312676,0.9834057
+above,0.9946324,0.9943074,1.0000002,0.9992869,0.99030274,0.9904754
+under,0.9929318,0.99233896,0.9992869,1.0,0.98967177,0.9901939
+far,0.98346335,0.98312676,0.99030274,0.98967177,1.0000002,0.9988357
+close,0.9839893,0.9834057,0.9904754,0.9901939,0.9988357,1.0

correct_filter/results/nvila/correct_only/csv/similarity_800k_L15.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.99999994,0.9955255,0.9616093,0.9634563,0.9570518,0.9555799
+right,0.9955255,1.0,0.9596715,0.9647523,0.956402,0.9524902
+above,0.9616093,0.9596715,1.0000004,0.98979074,0.9572837,0.95708257
+under,0.9634563,0.9647523,0.98979074,1.0,0.957764,0.95645
+far,0.9570518,0.956402,0.9572837,0.957764,1.0000007,0.99549437
+close,0.9555799,0.9524902,0.95708257,0.95645,0.99549437,0.99999964

correct_filter/results/nvila/correct_only/csv/similarity_800k_L16.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999993,0.937419,0.91674423,0.91465557,0.9169775,0.923213
+right,0.937419,0.9999999,0.9187542,0.9189402,0.92452234,0.9273007
+above,0.91674423,0.9187542,1.0,0.89840555,0.92467207,0.92687386
+under,0.91465557,0.9189402,0.89840555,1.0000001,0.90919626,0.9162538
+far,0.9169775,0.92452234,0.92467207,0.90919626,1.0000004,0.9890829
+close,0.923213,0.9273007,0.92687386,0.9162538,0.9890829,0.99999994

correct_filter/results/nvila/correct_only/csv/similarity_800k_L18.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000001,0.88733137,0.8091318,0.83194005,0.8167282,0.81322694
+right,0.88733137,1.0000002,0.83587855,0.8211735,0.8257367,0.8230107
+above,0.8091318,0.83587855,0.9999997,0.7676464,0.8233376,0.7811038
+under,0.83194005,0.8211735,0.7676464,0.99999976,0.81050754,0.8206099
+far,0.8167282,0.8257367,0.8233376,0.81050754,1.0000004,0.94082886
+close,0.81322694,0.8230107,0.7811038,0.8206099,0.94082886,1.0000004

correct_filter/results/nvila/correct_only/csv/similarity_800k_L19.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000002,0.904862,0.80710673,0.8477925,0.83130246,0.8128334
+right,0.904862,1.0000001,0.83244157,0.8415096,0.837732,0.8282201
+above,0.80710673,0.83244157,1.0,0.7711587,0.799425,0.7575373
+under,0.8477925,0.8415096,0.7711587,0.9999999,0.81178087,0.8155661
+far,0.83130246,0.837732,0.799425,0.81178087,0.99999976,0.940137
+close,0.8128334,0.8282201,0.7575373,0.8155661,0.940137,1.0

correct_filter/results/nvila/correct_only/csv/similarity_800k_L20.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.99999976,0.9131943,0.83245647,0.8717643,0.8463599,0.8216339
+right,0.9131943,1.0000001,0.84933776,0.87377536,0.85917556,0.84626216
+above,0.83245647,0.84933776,1.0,0.8021531,0.81296337,0.77681535
+under,0.8717643,0.87377536,0.8021531,0.9999999,0.8474405,0.84407175
+far,0.8463599,0.85917556,0.81296337,0.8474405,0.9999999,0.958832
+close,0.8216339,0.84626216,0.77681535,0.84407175,0.958832,0.9999998

correct_filter/results/nvila/correct_only/csv/similarity_800k_L27.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.99999976,0.9989176,0.99486405,0.9962475,0.9953634,0.99469465
+right,0.9989176,1.0000006,0.99491745,0.9964434,0.9955484,0.99498737
+above,0.99486405,0.99491745,1.0000001,0.9970196,0.9946532,0.99397403
+under,0.9962475,0.9964434,0.9970196,1.0000004,0.9958383,0.99531734
+far,0.9953634,0.9955484,0.9946532,0.9958383,1.0000002,0.9996456
+close,0.99469465,0.99498737,0.99397403,0.99531734,0.9996456,1.0000005

correct_filter/results/nvila/correct_only/csv/similarity_800k_L4.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999999,0.99931574,0.9958405,0.9904083,0.99077135,0.98957026
+right,0.99931574,1.0000001,0.99584186,0.99013525,0.9912202,0.98926467
+above,0.9958405,0.99584186,1.0000002,0.997966,0.99544257,0.994243
+under,0.9904083,0.99013525,0.997966,1.0,0.99456835,0.99469686
+far,0.99077135,0.9912202,0.99544257,0.99456835,1.0000001,0.9977426
+close,0.98957026,0.98926467,0.994243,0.99469686,0.9977426,1.0000002

correct_filter/results/nvila/correct_only/csv/similarity_800k_L7.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999998,0.99970233,0.9979286,0.9962451,0.99490523,0.9946172
+right,0.99970233,0.9999998,0.9979725,0.9960514,0.99498373,0.9943552
+above,0.9979286,0.9979725,0.99999994,0.99914914,0.9973988,0.9966422
+under,0.9962451,0.9960514,0.99914914,1.0000001,0.9971471,0.9969803
+far,0.99490523,0.99498373,0.9973988,0.9971471,1.0000006,0.99878615
+close,0.9946172,0.9943552,0.9966422,0.9969803,0.99878615,0.9999996

correct_filter/results/nvila/correct_only/csv/similarity_800k_L8.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000002,0.9997558,0.99791306,0.99646586,0.99523884,0.9950807
+right,0.9997558,1.0000001,0.9978681,0.99624777,0.99530035,0.99486065
+above,0.99791306,0.9978681,0.9999996,0.9992501,0.99727607,0.9967006
+under,0.99646586,0.99624777,0.9992501,0.9999994,0.99712163,0.99706066
+far,0.99523884,0.99530035,0.99727607,0.99712163,0.9999997,0.99886113
+close,0.9950807,0.99486065,0.9967006,0.99706066,0.99886113,1.0

correct_filter/results/nvila/correct_only/csv/similarity_80k_L0.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000002,0.9999928,0.9996275,0.9996228,0.9995172,0.9995024
+right,0.9999928,0.9999999,0.9996372,0.9996381,0.9995251,0.9995048
+above,0.9996275,0.9996372,1.0000001,0.9999921,0.9996672,0.9996296
+under,0.9996228,0.9996381,0.9999921,0.9999999,0.99967223,0.9996285
+far,0.9995172,0.9995251,0.9996672,0.99967223,1.0,0.9999763
+close,0.9995024,0.9995048,0.9996296,0.9996285,0.9999763,0.99999976

correct_filter/results/nvila/correct_only/csv/similarity_80k_L1.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0,0.9999891,0.9995492,0.9995388,0.99923646,0.99923515
+right,0.9999891,1.0,0.999559,0.9995541,0.9992583,0.999249
+above,0.9995492,0.999559,1.0,0.99998856,0.9993894,0.9993697
+under,0.9995388,0.9995541,0.99998856,1.0,0.9994066,0.9993794
+far,0.99923646,0.9992583,0.9993894,0.9994066,0.99999994,0.9999724
+close,0.99923515,0.999249,0.9993697,0.9993794,0.9999724,0.9999996

correct_filter/results/nvila/correct_only/csv/similarity_80k_L12.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000001,0.99992007,0.98682374,0.98694927,0.96131676,0.96126896
+right,0.99992007,1.0000005,0.98669213,0.9868369,0.9610707,0.96103275
+above,0.98682374,0.98669213,1.0000001,0.9996522,0.97436637,0.97421056
+under,0.98694927,0.9868369,0.9996522,0.99999976,0.9743051,0.9741693
+far,0.96131676,0.9610707,0.97436637,0.9743051,1.0000004,0.99982387
+close,0.96126896,0.96103275,0.97421056,0.9741693,0.99982387,1.0000001

correct_filter/results/nvila/correct_only/csv/similarity_80k_L13.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000002,0.99964416,0.9821861,0.9828103,0.9578027,0.95762783
+right,0.99964416,1.0000005,0.98217076,0.98289746,0.95772654,0.95757335
+above,0.9821861,0.98217076,0.99999976,0.9989925,0.97151095,0.9707316
+under,0.9828103,0.98289746,0.9989925,1.0000004,0.97208655,0.9714904
+far,0.9578027,0.95772654,0.97151095,0.97208655,0.9999999,0.9997549
+close,0.95762783,0.95757335,0.9707316,0.9714904,0.9997549,0.99999946

correct_filter/results/nvila/correct_only/csv/similarity_80k_L18.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999999,0.9506794,0.88382566,0.88530993,0.8269373,0.8271559
+right,0.9506794,1.0000001,0.87476265,0.90188134,0.838693,0.84223974
+above,0.88382566,0.87476265,1.0000002,0.9400799,0.9003339,0.8796129
+under,0.88530993,0.90188134,0.9400799,1.0000001,0.8822267,0.894028
+far,0.8269373,0.838693,0.9003339,0.8822267,0.9999995,0.99074644
+close,0.8271559,0.84223974,0.8796129,0.894028,0.99074644,1.0000006

correct_filter/results/nvila/correct_only/csv/similarity_80k_L2.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999998,0.99997544,0.99910533,0.99907017,0.99487907,0.99464613
+right,0.99997544,0.9999998,0.99910426,0.9991061,0.9950284,0.9947916
+above,0.99910533,0.99910426,0.99999946,0.9999348,0.994704,0.994506
+under,0.99907017,0.9991061,0.9999348,0.99999976,0.9948836,0.99467635
+far,0.99487907,0.9950284,0.994704,0.9948836,1.0000004,0.9999575
+close,0.99464613,0.9947916,0.994506,0.99467635,0.9999575,1.0000004

correct_filter/results/nvila/correct_only/csv/similarity_80k_L20.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000002,0.94373393,0.890359,0.8972434,0.8310231,0.83172584
+right,0.94373393,0.9999999,0.8833537,0.9212339,0.85914224,0.86019874
+above,0.890359,0.8833537,0.99999994,0.9401983,0.88969404,0.8716103
+under,0.8972434,0.9212339,0.9401983,1.0,0.89284706,0.9008564
+far,0.8310231,0.85914224,0.88969404,0.89284706,1.0000004,0.99279314
+close,0.83172584,0.86019874,0.8716103,0.9008564,0.99279314,1.0

correct_filter/results/nvila/correct_only/csv/similarity_80k_L24.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999996,0.97742325,0.88750535,0.8872726,0.8419494,0.8421845
+right,0.97742325,1.0000001,0.8867433,0.9016926,0.8560913,0.85670274
+above,0.88750535,0.8867433,1.0000001,0.971692,0.8757001,0.87120354
+under,0.8872726,0.9016926,0.971692,1.0000004,0.8741759,0.8779786
+far,0.8419494,0.8560913,0.8757001,0.8741759,0.9999998,0.9972954
+close,0.8421845,0.85670274,0.87120354,0.8779786,0.9972954,0.99999976

correct_filter/results/nvila/correct_only/csv/similarity_80k_L26.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999999,0.98567736,0.89949954,0.9015679,0.83705455,0.83549917
+right,0.98567736,0.9999999,0.89631855,0.90650165,0.84298885,0.841433
+above,0.89949954,0.89631855,0.99999976,0.97662956,0.8449816,0.840574
+under,0.9015679,0.90650165,0.97662956,1.0000002,0.84031403,0.84105325
+far,0.83705455,0.84298885,0.8449816,0.84031403,0.9999999,0.9974231
+close,0.83549917,0.841433,0.840574,0.84105325,0.9974231,1.0000002

correct_filter/results/nvila/correct_only/csv/similarity_80k_L5.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0,0.99994904,0.99607235,0.99601716,0.9891803,0.98881525
+right,0.99994904,0.9999997,0.99618846,0.9961636,0.9892961,0.98891604
+above,0.99607235,0.99618846,0.99999994,0.9999412,0.99260557,0.99223334
+under,0.99601716,0.9961636,0.9999412,1.0,0.992865,0.9924812
+far,0.9891803,0.9892961,0.99260557,0.992865,0.9999999,0.9999319
+close,0.98881525,0.98891604,0.99223334,0.9924812,0.9999319,1.0

correct_filter/results/nvila/correct_only/csv/similarity_80k_L7.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000007,0.99997765,0.9980402,0.99805593,0.9940974,0.99391335
+right,0.99997765,0.9999998,0.9980756,0.9981079,0.99421495,0.9940232
+above,0.9980402,0.9980756,1.0000001,0.9999677,0.9959363,0.99573493
+under,0.99805593,0.9981079,0.9999677,1.0000001,0.9960644,0.9958569
+far,0.9940974,0.99421495,0.9959363,0.9960644,0.99999994,0.99996877
+close,0.99391335,0.9940232,0.99573493,0.9958569,0.99996877,1.0000002

correct_filter/results/nvila/correct_only/csv/similarity_80k_L9.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,0.9999999,0.9999779,0.9979099,0.9978664,0.9914892,0.99141896
+right,0.9999779,0.99999994,0.99795455,0.9979277,0.9915622,0.9914869
+above,0.9979099,0.99795455,0.99999994,0.9999599,0.9929552,0.99283284
+under,0.9978664,0.9979277,0.9999599,0.9999999,0.99296045,0.99283123
+far,0.9914892,0.9915622,0.9929552,0.99296045,1.0,0.9999687
+close,0.99141896,0.9914869,0.99283284,0.99283123,0.9999687,1.0000004

correct_filter/results/nvila/correct_only/csv/similarity_roborefer_L12.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000004,0.9998851,0.9772084,0.9767925,0.92733586,0.92611945
+right,0.9998851,1.0,0.9774005,0.97699416,0.9275956,0.9263429
+above,0.9772084,0.9774005,0.9999999,0.9994277,0.9484417,0.9475716
+under,0.9767925,0.97699416,0.9994277,0.9999995,0.94778425,0.9468689
+far,0.92733586,0.9275956,0.9484417,0.94778425,0.99999994,0.9996827
+close,0.92611945,0.9263429,0.9475716,0.9468689,0.9996827,1.0000001

correct_filter/results/nvila/correct_only/csv/similarity_roborefer_L13.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0000001,0.999737,0.9714929,0.9713556,0.9228223,0.92111427
+right,0.999737,0.99999946,0.9718181,0.9717809,0.9234044,0.9216598
+above,0.9714929,0.9718181,1.0000006,0.99914813,0.9459721,0.9443468
+under,0.9713556,0.9717809,0.99914813,0.9999998,0.9453688,0.9437411
+far,0.9228223,0.9234044,0.9459721,0.9453688,0.99999994,0.99955255
+close,0.92111427,0.9216598,0.9443468,0.9437411,0.99955255,0.9999999

correct_filter/results/nvila/correct_only/csv/similarity_roborefer_L16.csv

@@ -0,0 +1,7 @@
+,left,right,above,under,far,close
+left,1.0,0.9601641,0.95128024,0.9552473,0.9148139,0.9127438
+right,0.9601641,1.0,0.947104,0.9543113,0.9126753,0.91043645
+above,0.95128024,0.947104,1.0000002,0.98705524,0.9433115,0.9390137
+under,0.9552473,0.9543113,0.98705524,1.0000002,0.9456221,0.9441068
+far,0.9148139,0.9126753,0.9433115,0.9456221,1.0000004,0.99866474
+close,0.9127438,0.91043645,0.9390137,0.9441068,0.99866474,1.0