| """Base pass rates of candidate Countdown difficulty settings (picks the generator settings of the puzzle pair). |
| |
| python -m tdprm.scripts.probe_countdown --backbone qwen1.5b [--per-config 160] [--n 8] |
| |
| Writes $TDPRM_HOME/report/probe_countdown_<backbone>.json: pass@1 and the share of prompts solved 1-7 of 8 times for |
| every (numbers, max number, max target) setting. |
| """ |
|
|
| import argparse |
| import json |
| import random |
|
|
| from .. import paths |
| from ..data.puzzles import countdown_puzzle, countdown_question |
| from ..data.verify import verify |
| from ..rl.launch import DOMAINS |
| from ..sampling.base_rollouts import model_path |
| from ..sampling.generate import generate, load_llm, render_prompt_ids |
|
|
| CONFIGS = [(3, 10, 100), (3, 15, 150), (3, 20, 200), (3, 30, 1000), (4, 10, 100), (4, 15, 200), (4, 25, 500), |
| (4, 50, 1000), (5, 10, 150), (5, 20, 500)] |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--backbone", default="qwen1.5b") |
| ap.add_argument("--per-config", type=int, default=160) |
| ap.add_argument("--n", type=int, default=8) |
| args = ap.parse_args() |
| rng = random.Random(7) |
| records = [] |
| for n, hi, top in CONFIGS: |
| k = 0 |
| while k < args.per_config: |
| got = countdown_puzzle(n, hi, rng, max_target=top, min_target=2) |
| if got is None: |
| continue |
| nums, target, expr = got |
| records.append({"domain": "countdown", "data_source": "countdown", "question": countdown_question(nums, target), |
| "answer": json.dumps({"answer": expr, "numbers": nums, "target": target}), |
| "config": f"{n}x{hi}<={top}"}) |
| k += 1 |
| from transformers import AutoTokenizer |
| path = model_path(args.backbone) |
| tok = AutoTokenizer.from_pretrained(path) |
| llm = load_llm(path, 2048, gpu_memory_utilization=0.85) |
| outs = generate(llm, [render_prompt_ids(tok, r) for r in records], args.n, 1.0, 1.0, |
| DOMAINS["countdown"]["max_response"], 0) |
| stats = {} |
| for r, samples in zip(records, outs): |
| c = sum(verify("countdown", text, r["answer"])["score"] for _, text, _ in samples) |
| s = stats.setdefault(r["config"], {"prompts": 0, "correct": 0, "kept": 0}) |
| s["prompts"] += 1 |
| s["correct"] += c |
| s["kept"] += 1 <= c <= args.n - 1 |
| for cfg, s in stats.items(): |
| s["pass@1"] = round(s.pop("correct") / (s["prompts"] * args.n), 3) |
| s["kept_share"] = round(s.pop("kept") / s["prompts"], 3) |
| out = paths.HOME / "report" / f"probe_countdown_{args.backbone}.json" |
| out.parent.mkdir(parents=True, exist_ok=True) |
| out.write_text(json.dumps(stats, indent=1)) |
| print(json.dumps(stats, indent=1)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
