phyground-code / dataprocessing /analysis /ablation_prompt_enhancement.py
anonymouscla's picture
Initial anonymous release: phyground-code
4949db9 verified
#!/usr/bin/env python3
"""
Ablation study: Effect of prompt enhancement (adding physical law/phenomenon descriptions)
on VLM evaluation scores.
Compares backup (pre-enhancement) eval files with current (post-enhancement) eval files,
analyzing score deltas per video where the prompt actually changed.
Run from the anonymous root dir:
python -m dataprocessing.analysis.ablation_prompt_enhancement
"""
import json
import os
import re
import sys
from collections import defaultdict
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
# ---------------------------------------------------------------------------
# Configuration
# ---------------------------------------------------------------------------
DATA_ROOT = "data"
BACKUP_DIR = os.path.join(DATA_ROOT, "backup_before_laws_update")
VIDEOS_DIR = os.path.join(DATA_ROOT, "videos")
GENERAL_METRICS = ["SA", "PTV", "persistence"]
SCORE_BINS = [(0, 1), (1, 2), (2, 3), (3, 4), (4, 5)]
# Law-to-domain mapping
LAW_TO_DOMAIN = {
# collision
"collision": "collision",
"impenetrability": "collision",
"momentum_transfer": "collision",
"momentum": "collision",
"elastic_deformation": "collision",
# gravity
"gravity": "gravity",
"free_fall": "gravity",
"projectile_motion": "gravity",
"buoyancy": "gravity",
# fluid
"fluid_continuity": "fluid",
"flow_dynamics": "fluid",
"flow": "fluid",
"viscosity": "fluid",
"surface_tension": "fluid",
"pressure": "fluid",
"continuity": "fluid",
# temporal / motion
"inertia": "temporal",
"acceleration": "temporal",
"velocity": "temporal",
"displacement": "temporal",
# lighting
"reflection": "lighting",
"refraction": "lighting",
"light_absorption": "lighting",
"shadow": "lighting",
"illumination": "lighting",
# deformation
"deformation": "deformation",
"plastic_deformation": "deformation",
# material
"material": "material",
"rigidity": "material",
"elasticity": "material",
"phase_transition": "material",
"melting": "material",
"combustion": "material",
}
# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------
def parse_judge_key(judge_str: str) -> str:
"""Extract a short judge key like 'gemini', 'qwen', 'gpt' from judge field."""
if not judge_str:
return "unknown"
return judge_str.split(":")[0]
def parse_filename_judge(filename: str) -> str:
"""Extract judge key from filename like eval_gemini_..., eval_qwen_permetric_..."""
m = re.match(r"eval_(gemini|qwen|gpt)", filename)
return m.group(1) if m else "unknown"
def is_permetric(filename: str) -> bool:
return "permetric" in filename
def get_physical_avg(result: dict) -> Optional[float]:
"""Extract physical average from a result entry."""
phys = result.get("physical")
if phys is None:
return None
if isinstance(phys, (int, float)):
return float(phys)
if isinstance(phys, dict):
avg = phys.get("avg")
if avg is not None:
return float(avg)
return None
def get_per_law_scores(result: dict) -> Dict[str, float]:
"""Extract per-law scores from a result entry."""
phys = result.get("physical")
if not isinstance(phys, dict):
return {}
laws = phys.get("laws", {})
out = {}
for law_name, law_data in laws.items():
if isinstance(law_data, dict) and "score" in law_data and law_data["score"] is not None:
out[law_name] = float(law_data["score"])
elif isinstance(law_data, (int, float)) and law_data is not None:
out[law_name] = float(law_data)
return out
def extract_dataset_from_video_dir(video_dir: str) -> str:
"""Extract dataset name from video_dir field, e.g. 'data/videos/cosmos-predict2.5-2b-wmb/' -> 'wmb'."""
vd = video_dir.rstrip("/")
basename = os.path.basename(vd)
# Try to find the dataset suffix: wmb, video_phy_2, physics_iq, openvid, video_phy_2
for ds in ["wmb", "video_phy_2", "physics_iq", "openvid", "video_phy_2", "wmb"]:
if basename.endswith(f"-{ds}"):
return ds
return basename
def extract_model_from_video_dir(video_dir: str) -> str:
"""Extract model name from video_dir."""
vd = video_dir.rstrip("/")
basename = os.path.basename(vd)
for ds in ["wmb", "video_phy_2", "physics_iq", "openvid", "video_phy_2", "wmb"]:
suffix = f"-{ds}"
if basename.endswith(suffix):
return basename[: -len(suffix)]
return basename
def load_eval_file(filepath: str) -> Optional[dict]:
"""Load an eval JSON file, returning None on error."""
try:
with open(filepath, "r") as f:
data = json.load(f)
if not data.get("results"):
return None
return data
except (json.JSONDecodeError, FileNotFoundError, KeyError):
return None
def get_timestamp_from_filename(filename: str) -> str:
"""Extract timestamp from filename like eval_gemini_20260322_200226.json."""
m = re.search(r"(\d{8}_\d{6})", filename)
return m.group(1) if m else ""
# ---------------------------------------------------------------------------
# Core: find old/new pairs
# ---------------------------------------------------------------------------
def find_pairs() -> List[Dict[str, Any]]:
"""
Find all (old_file, new_file) pairs for comparison.
Returns list of dicts with keys: old_path, new_path, model, dataset, judge, mode (batched/permetric).
"""
pairs = []
if not os.path.isdir(BACKUP_DIR):
print(f"ERROR: Backup directory not found: {BACKUP_DIR}", file=sys.stderr)
return pairs
for backup_model in sorted(os.listdir(BACKUP_DIR)):
backup_model_dir = os.path.join(BACKUP_DIR, backup_model)
if not os.path.isdir(backup_model_dir):
continue
for fname in sorted(os.listdir(backup_model_dir)):
if not fname.startswith("eval_") or not fname.endswith(".json"):
continue
old_path = os.path.join(backup_model_dir, fname)
old_data = load_eval_file(old_path)
if old_data is None:
continue
video_dir = old_data.get("video_dir", "")
if not video_dir:
continue
dataset = extract_dataset_from_video_dir(video_dir)
model = extract_model_from_video_dir(video_dir)
judge = parse_judge_key(old_data.get("judge", ""))
mode = "permetric" if is_permetric(fname) else "batched"
# Find corresponding current directory
current_dir = os.path.join(VIDEOS_DIR, f"{model}-{dataset}")
if not os.path.isdir(current_dir):
continue
# Find newest matching eval file in current dir
old_timestamp = get_timestamp_from_filename(fname)
best_new_path = None
best_new_ts = ""
for cur_fname in os.listdir(current_dir):
if not cur_fname.startswith("eval_") or not cur_fname.endswith(".json"):
continue
if ".old_pre_t26." in cur_fname:
continue
cur_judge = parse_filename_judge(cur_fname)
cur_mode = "permetric" if is_permetric(cur_fname) else "batched"
if cur_judge != judge or cur_mode != mode:
continue
cur_ts = get_timestamp_from_filename(cur_fname)
# Skip files with same timestamp as the old file (these are copies)
if cur_ts == old_timestamp:
continue
# Must be newer than old
if cur_ts <= old_timestamp:
continue
# Pick the newest one
if cur_ts > best_new_ts:
best_new_ts = cur_ts
best_new_path = os.path.join(current_dir, cur_fname)
if best_new_path:
pairs.append(
{
"old_path": old_path,
"new_path": best_new_path,
"model": model,
"dataset": dataset,
"judge": judge,
"mode": mode,
}
)
return pairs
# ---------------------------------------------------------------------------
# Core: compute deltas
# ---------------------------------------------------------------------------
def compute_deltas(
old_data: dict, new_data: dict
) -> List[Dict[str, Any]]:
"""
For each video present in both old and new where the prompt changed,
compute score deltas.
Returns list of per-video delta records.
"""
old_by_video = {r["video"]: r for r in old_data["results"]}
new_by_video = {r["video"]: r for r in new_data["results"]}
deltas = []
common_videos = set(old_by_video.keys()) & set(new_by_video.keys())
for vid in sorted(common_videos):
old_r = old_by_video[vid]
new_r = new_by_video[vid]
# Only analyze videos where the prompt actually changed
if old_r.get("prompt", "") == new_r.get("prompt", ""):
continue
rec: Dict[str, Any] = {
"video": vid,
"old_prompt": old_r.get("prompt", ""),
"new_prompt": new_r.get("prompt", ""),
"physical_laws": new_r.get("physical_laws", old_r.get("physical_laws", [])),
}
# General metrics deltas
for m in GENERAL_METRICS:
old_val = old_r.get(m)
new_val = new_r.get(m)
if old_val is not None and new_val is not None:
rec[f"{m}_old"] = float(old_val)
rec[f"{m}_new"] = float(new_val)
rec[f"{m}_delta"] = float(new_val) - float(old_val)
# general_avg delta
old_ga = old_r.get("general_avg")
new_ga = new_r.get("general_avg")
if old_ga is not None and new_ga is not None:
rec["general_avg_old"] = float(old_ga)
rec["general_avg_new"] = float(new_ga)
rec["general_avg_delta"] = float(new_ga) - float(old_ga)
# physical_avg delta
old_pa = get_physical_avg(old_r)
new_pa = get_physical_avg(new_r)
if old_pa is not None and new_pa is not None:
rec["physical_avg_old"] = float(old_pa)
rec["physical_avg_new"] = float(new_pa)
rec["physical_avg_delta"] = float(new_pa) - float(old_pa)
# Per-law score deltas
old_laws = get_per_law_scores(old_r)
new_laws = get_per_law_scores(new_r)
law_deltas = {}
for law in set(old_laws.keys()) & set(new_laws.keys()):
law_deltas[law] = new_laws[law] - old_laws[law]
rec["per_law_deltas"] = law_deltas
deltas.append(rec)
return deltas
# ---------------------------------------------------------------------------
# Aggregation
# ---------------------------------------------------------------------------
def safe_mean(values: list) -> Optional[float]:
if not values:
return None
return sum(values) / len(values)
def format_delta(val: Optional[float], decimals: int = 4) -> str:
if val is None:
return "N/A"
sign = "+" if val >= 0 else ""
return f"{sign}{val:.{decimals}f}"
def format_float(val: Optional[float], decimals: int = 4) -> str:
if val is None:
return "N/A"
return f"{val:.{decimals}f}"
class AblationAnalysis:
def __init__(self):
# Each entry: (model, dataset, judge, mode, delta_record)
self.all_records: List[Tuple[str, str, str, str, Dict[str, Any]]] = []
def add(self, model: str, dataset: str, judge: str, mode: str, deltas: List[Dict[str, Any]]):
for d in deltas:
self.all_records.append((model, dataset, judge, mode, d))
def _filter(
self,
model: Optional[str] = None,
dataset: Optional[str] = None,
judge: Optional[str] = None,
) -> List[Dict[str, Any]]:
out = []
for m, ds, j, mode, rec in self.all_records:
if model and m != model:
continue
if dataset and ds != dataset:
continue
if judge and j != judge:
continue
out.append(rec)
return out
def _metric_deltas(self, records: List[Dict], metric_key: str) -> List[float]:
key = f"{metric_key}_delta"
return [r[key] for r in records if key in r]
def overall_summary(self) -> str:
lines = []
records = self._filter()
n = len(records)
lines.append(f"**Total video comparisons (prompt changed):** {n}")
lines.append("")
if n == 0:
lines.append("No data to analyze.")
return "\n".join(lines)
# Table header
metrics = GENERAL_METRICS + ["general_avg", "physical_avg"]
lines.append("| Metric | Mean Delta | Median Delta | Std Dev | N |")
lines.append("|--------|-----------|-------------|---------|---|")
for metric in metrics:
vals = self._metric_deltas(records, metric)
if not vals:
lines.append(f"| {metric} | N/A | N/A | N/A | 0 |")
continue
import statistics
mean = statistics.mean(vals)
median = statistics.median(vals)
stdev = statistics.stdev(vals) if len(vals) > 1 else 0.0
lines.append(
f"| {metric} | {format_delta(mean)} | {format_delta(median)} | {format_float(stdev)} | {len(vals)} |"
)
return "\n".join(lines)
def per_group_table(self, group_key: str) -> str:
"""Group by model, dataset, or judge."""
groups: Dict[str, List[Dict]] = defaultdict(list)
for m, ds, j, mode, rec in self.all_records:
if group_key == "model":
key = m
elif group_key == "dataset":
key = ds
elif group_key == "judge":
key = j
else:
key = "all"
groups[key].append(rec)
metrics = GENERAL_METRICS + ["general_avg", "physical_avg"]
lines = []
header = f"| {group_key.capitalize()} | N |"
sep = "|---|---|"
for metric in metrics:
header += f" {metric} |"
sep += "---|"
lines.append(header)
lines.append(sep)
for gname in sorted(groups.keys()):
recs = groups[gname]
row = f"| {gname} | {len(recs)} |"
for metric in metrics:
vals = self._metric_deltas(recs, metric)
mean = safe_mean(vals)
row += f" {format_delta(mean)} |"
lines.append(row)
return "\n".join(lines)
def per_domain_table(self) -> str:
"""Aggregate per-law deltas into domains."""
domain_deltas: Dict[str, List[float]] = defaultdict(list)
for _, _, _, _, rec in self.all_records:
for law, delta in rec.get("per_law_deltas", {}).items():
domain = LAW_TO_DOMAIN.get(law, "other")
domain_deltas[domain].append(delta)
lines = []
lines.append("| Domain | Mean Delta | N | Improved% | Degraded% | Unchanged% |")
lines.append("|--------|-----------|---|----------|----------|-----------|")
for domain in sorted(domain_deltas.keys()):
vals = domain_deltas[domain]
n = len(vals)
mean = safe_mean(vals)
improved = sum(1 for v in vals if v > 0) / n * 100 if n else 0
degraded = sum(1 for v in vals if v < 0) / n * 100 if n else 0
unchanged = sum(1 for v in vals if v == 0) / n * 100 if n else 0
lines.append(
f"| {domain} | {format_delta(mean)} | {n} | {improved:.1f}% | {degraded:.1f}% | {unchanged:.1f}% |"
)
return "\n".join(lines)
def per_law_table(self) -> str:
"""Per-law breakdown."""
law_deltas: Dict[str, List[float]] = defaultdict(list)
for _, _, _, _, rec in self.all_records:
for law, delta in rec.get("per_law_deltas", {}).items():
law_deltas[law].append(delta)
lines = []
lines.append("| Law | Domain | Mean Delta | N | Improved% | Degraded% |")
lines.append("|-----|--------|-----------|---|----------|----------|")
# Sort by mean delta descending
sorted_laws = sorted(
law_deltas.keys(), key=lambda l: safe_mean(law_deltas[l]) or 0, reverse=True
)
for law in sorted_laws:
vals = law_deltas[law]
n = len(vals)
mean = safe_mean(vals)
domain = LAW_TO_DOMAIN.get(law, "other")
improved = sum(1 for v in vals if v > 0) / n * 100 if n else 0
degraded = sum(1 for v in vals if v < 0) / n * 100 if n else 0
lines.append(
f"| {law} | {domain} | {format_delta(mean)} | {n} | {improved:.1f}% | {degraded:.1f}% |"
)
return "\n".join(lines)
def score_bin_table(self) -> str:
"""Effect on low-score vs high-score videos, binned by old general_avg."""
bin_deltas: Dict[str, Dict[str, List[float]]] = {}
for lo, hi in SCORE_BINS:
label = f"{lo}-{hi}"
bin_deltas[label] = defaultdict(list)
metrics = GENERAL_METRICS + ["general_avg", "physical_avg"]
for _, _, _, _, rec in self.all_records:
old_ga = rec.get("general_avg_old")
if old_ga is None:
continue
for lo, hi in SCORE_BINS:
if lo <= old_ga < hi or (hi == 5 and old_ga == 5):
label = f"{lo}-{hi}"
for metric in metrics:
key = f"{metric}_delta"
if key in rec:
bin_deltas[label][metric].append(rec[key])
break
lines = []
header = "| Old general_avg Bin | N |"
sep = "|---|---|"
for metric in metrics:
header += f" {metric} |"
sep += "---|"
lines.append(header)
lines.append(sep)
for lo, hi in SCORE_BINS:
label = f"{lo}-{hi}"
bd = bin_deltas[label]
# N = number of records in this bin
n_vals = bd.get("general_avg", [])
n = len(n_vals)
row = f"| {label} | {n} |"
for metric in metrics:
vals = bd.get(metric, [])
mean = safe_mean(vals)
row += f" {format_delta(mean)} |"
lines.append(row)
return "\n".join(lines)
def prompt_length_analysis(self) -> str:
"""Analyze whether longer prompt additions correlate with bigger deltas."""
records_with_len = []
for _, _, _, _, rec in self.all_records:
old_len = len(rec.get("old_prompt", ""))
new_len = len(rec.get("new_prompt", ""))
delta_len = new_len - old_len
ga_delta = rec.get("general_avg_delta")
pa_delta = rec.get("physical_avg_delta")
if ga_delta is not None:
records_with_len.append((delta_len, ga_delta, pa_delta))
if not records_with_len:
return "No data for prompt length analysis."
lines = []
# Bin by prompt length increase
bins = [(0, 30), (30, 60), (60, 90), (90, 200)]
lines.append("| Prompt Length Increase | N | Mean general_avg Delta | Mean physical_avg Delta |")
lines.append("|----------------------|---|----------------------|----------------------|")
for lo, hi in bins:
subset = [(d, g, p) for d, g, p in records_with_len if lo <= d < hi]
n = len(subset)
ga_mean = safe_mean([g for _, g, _ in subset])
pa_mean = safe_mean([p for _, _, p in subset if p is not None])
lines.append(
f"| {lo}-{hi} chars | {n} | {format_delta(ga_mean)} | {format_delta(pa_mean)} |"
)
return "\n".join(lines)
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
def main():
# Ensure we're in anonymous root or adjust paths
if not os.path.isdir(DATA_ROOT):
# Try from the script's location
script_dir = os.path.dirname(os.path.abspath(__file__))
anonymous_root = os.path.abspath(os.path.join(script_dir, "..", ".."))
os.chdir(anonymous_root)
if not os.path.isdir(DATA_ROOT):
print(f"ERROR: Cannot find {DATA_ROOT}. Run from the anonymous root directory.", file=sys.stderr)
sys.exit(1)
print("# Ablation Study: Effect of Prompt Enhancement on VLM Evaluation Scores")
print()
print("Comparing backup (pre-enhancement) vs current (post-enhancement) evaluation files.")
print("Only videos where the prompt actually changed between versions are included.")
print()
# Step 1: Find all old/new pairs
pairs = find_pairs()
print(f"Found **{len(pairs)}** old/new eval file pairs.")
print()
if not pairs:
print("No pairs found. Exiting.")
return
# Print pairs summary
print("## File Pairs Found")
print()
print("| # | Model | Dataset | Judge | Mode | Old File | New File |")
print("|---|-------|---------|-------|------|----------|----------|")
for i, p in enumerate(pairs, 1):
old_base = os.path.basename(p["old_path"])
new_base = os.path.basename(p["new_path"])
print(
f"| {i} | {p['model']} | {p['dataset']} | {p['judge']} | {p['mode']} | {old_base} | {new_base} |"
)
print()
# Step 2: Compute deltas for each pair
analysis = AblationAnalysis()
total_videos = 0
total_changed = 0
skipped_pairs = 0
for p in pairs:
old_data = load_eval_file(p["old_path"])
new_data = load_eval_file(p["new_path"])
if old_data is None or new_data is None:
skipped_pairs += 1
continue
old_by_video = {r["video"]: r for r in old_data["results"]}
new_by_video = {r["video"]: r for r in new_data["results"]}
common = set(old_by_video.keys()) & set(new_by_video.keys())
total_videos += len(common)
deltas = compute_deltas(old_data, new_data)
total_changed += len(deltas)
analysis.add(p["model"], p["dataset"], p["judge"], p["mode"], deltas)
print(f"**Total matched videos across all pairs:** {total_videos}")
print(f"**Videos with prompt changes:** {total_changed}")
if skipped_pairs:
print(f"**Skipped pairs (load errors):** {skipped_pairs}")
print()
if total_changed == 0:
print("No videos with prompt changes found. Exiting.")
return
# Step 3: Reports
print("---")
print()
print("## 1. Overall Score Deltas (All Models, Datasets, Judges)")
print()
print(analysis.overall_summary())
print()
print("---")
print()
print("## 2. Per-Model Breakdown")
print()
print(analysis.per_group_table("model"))
print()
print("---")
print()
print("## 3. Per-Dataset Breakdown")
print()
print(analysis.per_group_table("dataset"))
print()
print("---")
print()
print("## 4. Per-Judge Breakdown")
print()
print(analysis.per_group_table("judge"))
print()
print("---")
print()
print("## 5. Per-Domain Breakdown (Physical Laws)")
print()
print(analysis.per_domain_table())
print()
print("---")
print()
print("## 6. Per-Law Breakdown (sorted by mean delta, descending)")
print()
print(analysis.per_law_table())
print()
print("---")
print()
print("## 7. Effect by Old Score Bin")
print()
print("Videos binned by their old general_avg score:")
print()
print(analysis.score_bin_table())
print()
print("---")
print()
print("## 8. Prompt Length Increase Analysis")
print()
print(analysis.prompt_length_analysis())
print()
if __name__ == "__main__":
main()