phyground-code / dataprocessing /refine /gen_hard_subset.py
anonymouscla's picture
Initial anonymous release: phyground-code
4949db9 verified
"""Generate Anonymous-Hard subset from Qwen per-law eval results.
For each prompt, computes a physics micro-avg from per-law scores
(all 13 laws), averaged across all models.
Keeps prompts with cross-model avg < threshold.
Physical laws are sourced from the canonical prompt JSONs (not from eval
JSONs, which may contain stale vocabulary).
Usage:
# Default: threshold 1.75, overwrite existing hard subset
python -m dataprocessing.refine.gen_hard_subset
# Custom threshold, write to a different file
python -m dataprocessing.refine.gen_hard_subset --threshold 1.50 \
--output data/prompts/anonymous_hard_subset_150.json
# Dry run: print stats without writing
python -m dataprocessing.refine.gen_hard_subset --dry-run
# Strict mode: fail on data quality issues
python -m dataprocessing.refine.gen_hard_subset --strict
"""
import argparse
import json
import logging
import sys
from collections import Counter, defaultdict
from pathlib import Path
from dataprocessing.common.pipeline import PipelineCheck
from dataprocessing.common.video_id import PROMPT_SOURCES, load_source_laws
logger = logging.getLogger(__name__)
ROOT = Path(__file__).resolve().parents[2]
VIDEOS_DIR = ROOT / "data/videos"
OUTPUT_PATH = ROOT / "data/prompts/anonymous_hard_subset.json"
# Dataset suffixes that appear in video dir names.
DATASET_SUFFIXES = ["video_phy_2", "physics_iq", "openvid", "wmb"]
def parse_model_dataset(dirname: str) -> tuple[str, str] | None:
"""Extract (model, dataset) from a directory name like 'ltx-2-video_phy_2'."""
for ds in DATASET_SUFFIXES:
if dirname.endswith(f"-{ds}"):
model = dirname[:-(len(ds) + 1)]
return model, ds
return None
def find_latest_eval(dirpath: Path, evaluator: str = "qwen") -> Path | None:
"""Find the latest batched eval JSON for the given evaluator, fallback to gemini."""
evals = sorted(dirpath.glob(f"eval_{evaluator}_2*.json"))
if evals:
return evals[-1]
if evaluator != "gemini":
gemini = sorted(dirpath.glob("eval_gemini*_2*.json"))
if gemini:
return gemini[-1]
return None
def load_eval_scores(eval_path: Path) -> list[dict]:
"""Load eval JSON and extract per-video physics micro-avg.
Computes micro-avg from per-law scores (all laws).
Follows the same scoring approach as score_histogram / rank.md.
Returns list of dicts with keys: video, prompt, phys_micro_avg, n_laws_scored.
"""
with open(eval_path) as f:
data = json.load(f)
entries = []
for r in data.get("results", []):
video = r.get("video", "")
prompt = r.get("prompt", "")
if not video:
continue
# Per-law physical scores (null-aware, supports v1 and v2 formats)
phys = r.get("physical", {})
if not isinstance(phys, dict):
continue
laws = phys.get("laws", {})
scored_vals = []
for law_name, law_data in laws.items():
if not isinstance(law_data, dict):
continue
score = law_data.get("score")
is_scored = (law_data.get("status") == "scored"
or law_data.get("valid", False))
if is_scored and score is not None:
scored_vals.append(score)
if not scored_vals:
continue
entries.append({
"video": video,
"prompt": prompt,
"phys_micro_avg": sum(scored_vals) / len(scored_vals),
"n_laws_scored": len(scored_vals),
})
return entries
def main(argv: list[str] | None = None):
parser = argparse.ArgumentParser(
description="Generate Anonymous-Hard subset from Gemini eval scores")
parser.add_argument("--threshold", type=float, default=3.00,
help="Physics micro-avg threshold (default: 3.00)")
parser.add_argument("--output", type=str, default=None,
help="Output path (default: data/prompts/anonymous_hard_subset.json)")
parser.add_argument("--dry-run", action="store_true",
help="Print stats without writing")
parser.add_argument("--strict", action="store_true",
help="Fail on data quality issues (for CI)")
args = parser.parse_args(argv)
logging.basicConfig(
level=logging.INFO,
format="[%(asctime)s] %(message)s",
handlers=[logging.StreamHandler(sys.stdout)],
)
checker = PipelineCheck(strict=args.strict)
# ---- Step 1: Collect per-model physics micro-avg for every video ----
# {eval_video: {model: phys_micro_avg}}
video_scores: dict[str, dict] = defaultdict(dict)
video_prompts: dict[str, str] = {}
video_eval_ds: dict[str, str] = {} # eval-side dataset suffix
eval_paths_by_ds: dict[str, Path] = {}
for d in sorted(VIDEOS_DIR.iterdir()):
if not d.is_dir():
continue
parsed = parse_model_dataset(d.name)
if parsed is None:
continue
model, dataset = parsed
if "real_world" in model:
continue
eval_path = find_latest_eval(d)
if eval_path is None:
continue
eval_paths_by_ds[dataset] = eval_path
entries = load_eval_scores(eval_path)
logger.info("Loaded %d entries from %s", len(entries), eval_path.name)
for e in entries:
vid = e["video"]
video_scores[vid][model] = e["phys_micro_avg"]
video_prompts[vid] = e["prompt"]
video_eval_ds[vid] = dataset
logger.info("Total unique videos with scores: %d", len(video_scores))
# ---- Step 2: Compute cross-model average ----
video_difficulty = {}
for vid, model_scores in video_scores.items():
vals = list(model_scores.values())
avg = sum(vals) / len(vals)
video_difficulty[vid] = {
"phys_micro_avg": round(avg, 3),
"n_models": len(model_scores),
}
# ---- Step 3: Filter by threshold ----
hard_vids = [
vid for vid, diff in video_difficulty.items()
if diff["phys_micro_avg"] < args.threshold
]
hard_vids.sort(key=lambda v: video_difficulty[v]["phys_micro_avg"])
logger.info("Threshold < %.2f: %d / %d videos",
args.threshold, len(hard_vids), len(video_difficulty))
# ---- Step 4: Look up physical_laws via canonical vid matching ----
source = load_source_laws()
# Check staleness: only compare each source JSON against its matching eval
DS_TO_EVAL_SUFFIX = {
"wmb": "wmb", "video_phy_2": "video_phy_2", "physics_iq": "physics_iq",
"openvid": "openvid",
}
for ds_name, src_path in PROMPT_SOURCES:
eval_suffix = DS_TO_EVAL_SUFFIX.get(ds_name)
if eval_suffix and eval_suffix in eval_paths_by_ds:
checker.check_staleness(src_path, eval_paths_by_ds[eval_suffix])
missing = 0
prompts_out = []
seen_disk_vids: dict[str, int] = {} # disk_vid -> index in prompts_out
for eval_vid in hard_vids:
eval_ds = video_eval_ds.get(eval_vid, "")
matched = source.resolve_eval(eval_vid, eval_ds)
if matched:
cvid, entry = matched
laws = entry["laws"]
dataset = entry["dataset"]
prompt = entry["prompt"] or video_prompts.get(eval_vid, "")
legacy_ids = source.cvid_to_legacies.get(cvid, set())
disk_vid = max(legacy_ids, key=len) if legacy_ids else eval_vid
else:
# Not in kept source — skip (removed, safety-blocked, etc.)
missing += 1
continue
# Deduplicate: different eval_vid values can resolve to the same
# disk_vid via legacy aliases. Keep the entry with more models.
if disk_vid in seen_disk_vids:
idx = seen_disk_vids[disk_vid]
existing = prompts_out[idx]
if video_difficulty[eval_vid]["n_models"] > existing["difficulty"]["n_models"]:
prompts_out[idx] = {
"video": disk_vid,
"dataset": dataset,
"prompt": prompt,
"physical_laws": laws,
"difficulty": video_difficulty[eval_vid],
"per_model_scores": dict(video_scores[eval_vid]),
}
continue
checker.check_empty_laws(disk_vid, laws, dataset,
resolved=matched is not None)
seen_disk_vids[disk_vid] = len(prompts_out)
prompts_out.append({
"video": disk_vid,
"dataset": dataset,
"prompt": prompt,
"physical_laws": laws,
"difficulty": video_difficulty[eval_vid],
"per_model_scores": dict(video_scores[eval_vid]),
})
checker.check_missing_ratio(missing, len(hard_vids))
# ---- Step 5: Compute stats and report ----
by_dataset = Counter(p["dataset"] for p in prompts_out)
law_counts = Counter()
for p in prompts_out:
for law in p["physical_laws"]:
law_counts[law] += 1
output = {
"description": (
f"Anonymous-Hard: prompts where cross-model physics micro-avg < {args.threshold} "
f"(Qwen, per-law scores, all 13 laws)"
),
"threshold": args.threshold,
"scoring_mode": "phys_micro_avg",
"judge": "qwen",
"num_prompts": len(prompts_out),
"by_dataset": dict(by_dataset.most_common()),
"prompts": prompts_out,
}
logger.info("=" * 60)
logger.info("Hard subset: %d prompts", len(prompts_out))
logger.info("By dataset:")
for ds, cnt in by_dataset.most_common():
logger.info(" %s: %d", ds, cnt)
logger.info("Physical law counts:")
for law, cnt in law_counts.most_common():
logger.info(" %s: %d", law, cnt)
score = checker.report()
if args.dry_run:
logger.info("(dry-run — no file written)")
checker.finalize()
return
out_path = Path(args.output) if args.output else OUTPUT_PATH
with open(out_path, "w") as f:
json.dump(output, f, indent=2, ensure_ascii=False)
logger.info("Saved → %s", out_path)
checker.finalize()
if __name__ == "__main__":
main()