| |
| """Rule-based grounding check: how often labels claim something the game's own |
| measured-feature card contradicts. Same rules for teachers and students. |
| |
| usage: grounding.py TRIAL [TRIAL ...] (samples_<TRIAL>.jsonl in results/labeler) |
| """ |
| import collections |
| import json |
| import os |
| import re |
| import sys |
|
|
| R = "/iopsstor/scratch/cscs/ibadanin/projects/llm_chess/results/labeler/" |
| VAL = os.environ.get("GROUND_SPLIT", "/iopsstor/scratch/cscs/ibadanin/projects/llm_chess/data/labeler_gemma/test.jsonl") |
| BANDS = ("terse_1", "terse_2", "short", "medium", "long") |
|
|
|
|
| def facts(u: str) -> dict: |
| m = re.search(r"sacrifices \(captures.*?\): (\d+)", u) |
| p = re.search(r"pressure on \w+'s king ring: [\d.]+ attackers/ply — (\w+)", u) |
| k = re.search(r"king: (\w+)", u) |
| d = re.search(r"development: ([\w ]+) \(", u) |
| return {"sac": int(m.group(1)) if m else None, "queens_traded": "queens: traded" in u, |
| "press": p.group(1) if p else None, "castle": k.group(1) if k else None, |
| "dev": d.group(1) if d else None} |
|
|
|
|
| RULES = [ |
| ("sacrifice", r"\bsacrific", lambda f: f["sac"] == 0), |
| ("queen trade", r"trade (the |your )?queens?|swap (the |your )?queens?|queens? off|queenless", |
| lambda f: not f["queens_traded"]), |
| ("king attack", r"attack (the|his|her|their|black.s|white.s|enemy|opponent.s) king|king ?hunt|storm the king|" |
| r"crash the king|mate attack", lambda f: f["press"] == "none"), |
| ("slow development", r"slow(ly)? (to )?develop|delay(ed)? development|leave pieces (at )?home", |
| lambda f: f["dev"] in ("very fast", "brisk")), |
| ("never castle", r"never castle|don.t castle|skip castling|leave (your )?king in the cent", |
| lambda f: f["castle"] in ("kingside", "queenside")), |
| ] |
|
|
|
|
| def score(label_sets): |
| """label_sets: list of (card_text, {band: text}). Returns (label rate, game rate, Counter).""" |
| c, n, bad = collections.Counter(), 0, 0 |
| for card, labs in label_sets: |
| f, hit = facts(card), False |
| for b in BANDS: |
| t = (labs.get(b) or "").lower() |
| n += 1 |
| for name, rx, cond in RULES: |
| if re.search(rx, t) and cond(f): |
| c[name] += 1 |
| hit = True |
| bad += hit |
| return sum(c.values()) / max(1, n), bad / max(1, len(label_sets)), c |
|
|
|
|
| def main(): |
| val = {json.loads(l)["key"]: json.loads(l) for l in open(VAL)} |
| trials = sys.argv[1:] |
| samples = {t: {json.loads(l)["key"]: json.loads(l) for l in open(R + f"samples_{t}.jsonl")} for t in trials} |
| keys = sorted(set.intersection(*[set(s) for s in samples.values()])) |
| rows = [("teacher (mix)", [(val[k]["user"], val[k]["labels"]) for k in keys]), |
| ("teacher=gemma", [(val[k]["user"], val[k]["labels"]) for k in keys if val[k]["labeler"] == "gemma"]), |
| ("teacher=apertus", [(val[k]["user"], val[k]["labels"]) for k in keys if val[k]["labeler"] == "apertus"])] |
| rows += [(t, [(val[k]["user"], samples[t][k]["pred"]) for k in keys]) for t in trials] |
| print(f"| source | games | labels contradicted | games with any | breakdown |\n|---|---|---|---|---|") |
| for name, sets in rows: |
| lr, gr, c = score(sets) |
| print(f"| {name} | {len(sets)} | {lr:.1%} | {gr:.0%} | " + ", ".join(f"{k} {v}" for k, v in c.most_common()) + " |") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
