| |
| |
| """Overlap of Udgam Logic Puzzles with public reasoning benchmarks that use similar puzzle types. |
| |
| python benchmark_overlap.py STAGE_DIR BENCH_DIR OUT.json |
| |
| BENCH_DIR holds local copies of lukaemon/bbh (all 27 tasks), K-and-K/knights-and-knaves (test) and |
| allenai/ZebraLogicBench (grid_mode, mc_mode). For each of our questions we look for (a) an identical normalised text and |
| (b) any shared run of 13 words (the GPT-3 style contamination check); 8-word runs are counted for information. |
| """ |
| import collections |
| import glob |
| import json |
| import os |
| import re |
| import sys |
|
|
| import pyarrow.parquet as pq |
|
|
|
|
| def words(t): |
| return re.findall(r"[a-z0-9]+", t.lower()) |
|
|
|
|
| def grams(ws, n): |
| return {" ".join(ws[i:i + n]) for i in range(len(ws) - n + 1)} if len(ws) >= n else {" ".join(ws)} |
|
|
|
|
| def bench_items(d): |
| for p in sorted(glob.glob(os.path.join(d, "bbh", "*", "*.parquet"))): |
| task = p.split(os.sep)[-2] |
| for r in pq.read_table(p).to_pylist(): |
| yield f"bbh/{task}", r["input"] |
| for p in sorted(glob.glob(os.path.join(d, "kk", "test", "*.jsonl"))): |
| for line in open(p): |
| yield "knights-and-knaves", json.loads(line)["quiz"] |
| for p in sorted(glob.glob(os.path.join(d, "zebra", "*", "*.parquet"))): |
| for r in pq.read_table(p).to_pylist(): |
| yield f"zebralogic/{p.split(os.sep)[-2]}", r["puzzle"] |
|
|
|
|
| def main(stage, bench, out): |
| ours = [] |
| for split in ("train", "test"): |
| ours += [(split, r["id"], r["family"], r["question"]) for r in pq.read_table(os.path.join(stage, "data", f"{split}.parquet")).to_pylist()] |
| exact, g13, g8, n_items = {}, collections.defaultdict(set), collections.defaultdict(set), collections.Counter() |
| for name, text in bench_items(bench): |
| n_items[name.split("/")[0]] += 1 |
| ws = words(text) |
| exact[" ".join(ws)] = name |
| for g in grams(ws, 13): |
| g13[g].add(name) |
| for g in grams(ws, 8): |
| g8[g].add(name) |
| hits13, hits8, hits_exact = collections.Counter(), collections.Counter(), 0 |
| examples = collections.defaultdict(list) |
| for split, pid, fam, q in ours: |
| ws = words(q) |
| if " ".join(ws) in exact: |
| hits_exact += 1 |
| shared = {g: g13[g] for g in grams(ws, 13) if g in g13} |
| if shared: |
| hits13[fam] += 1 |
| if len(examples[fam]) < 3: |
| g, src = next(iter(shared.items())) |
| examples[fam].append(dict(id=pid, benchmark=sorted(src)[0], shared_13_words=g)) |
| if any(g in g8 for g in grams(ws, 8)): |
| hits8[fam] += 1 |
| res = dict(benchmark_items=dict(n_items), our_questions=len(ours), exact_matches=hits_exact, |
| questions_sharing_13_words=dict(hits13), questions_sharing_8_words=dict(hits8), examples_13=examples) |
| json.dump(res, open(out, "w"), indent=1) |
| print(json.dumps(res, indent=1)[:3000]) |
|
|
|
|
| if __name__ == "__main__": |
| main(*sys.argv[1:4]) |
|
|
