| """Score Terminal-Bench 2.1 runs (~/tbench/jobs/tb21-<CFG>) and compare configs with paired statistics. |
| |
| Usage: python score_tb.py [JOBS_DIR] # default ~/tbench/jobs |
| A task passes when its verifier reward is 1.0 (the official metric). |
| """ |
| import json |
| import random |
| import sys |
| from pathlib import Path |
|
|
| J = Path(sys.argv[1]) if len(sys.argv) > 1 else Path.home() / "tbench" / "jobs" |
| CFGS = ["BF16", "BF16-FI", "FP8", "FP4"] |
| PAIRS = [("BF16-FI", "BF16", "kernel effect"), ("FP8", "BF16", "FP8 KV effect"), ("FP4", "BF16-FI", "NVFP4 KV effect")] |
|
|
|
|
| def load(cfg): |
| out = {} |
| for f in (J / f"tb21-{cfg}").glob("*__*/result.json"): |
| r = json.loads(f.read_text()) |
| rew = ((r.get("verifier_result") or {}).get("rewards") or {}).get("reward") or 0 |
| exc = (r.get("exception_info") or {}).get("exception_type") |
| out[r["task_name"].split("/")[-1]] = {"pass": 1 if rew >= 1 else 0, "exc": exc} |
| return out |
|
|
|
|
| def paired_ci(d, iters=10000, seed=0): |
| rng, n = random.Random(seed), len(d) |
| m = sorted(sum(d[rng.randrange(n)] for _ in range(n)) / n for _ in range(iters)) |
| return sum(d) / n, m[int(0.025 * iters)], m[int(0.975 * iters)] |
|
|
|
|
| R = {c: load(c) for c in CFGS} |
| print("Terminal-Bench 2.1 (89 tasks, terminus-2, reasoning_effort=xhigh, k=1, agent timeout x4)") |
| print(f"{'run':<10}{'done':>8}{'passed':>8}{'failed':>8}{'pass %':>9}{'timeouts':>10}{'other errs':>12}") |
| for c in CFGS: |
| v = R[c].values() |
| n, p = len(v), sum(x["pass"] for x in v) |
| to = sum(x["exc"] == "AgentTimeoutError" for x in v) |
| oe = sum(bool(x["exc"]) and x["exc"] != "AgentTimeoutError" for x in v) |
| print(f"{c:<10}{n:>5}/89{p:>8}{n - p:>8}{100 * p / max(n, 1):>9.1f}{to:>10}{oe:>12}") |
|
|
| print("\nPaired differences (treatment - baseline, points), 95% bootstrap CI, flips pass->fail / fail->pass") |
| for t, b, label in PAIRS: |
| ids = sorted(set(R[t]) & set(R[b])) |
| if not ids: |
| continue |
| m, lo, hi = paired_ci([R[t][i]["pass"] - R[b][i]["pass"] for i in ids]) |
| pf = sum(R[b][i]["pass"] > R[t][i]["pass"] for i in ids) |
| fp = sum(R[b][i]["pass"] < R[t][i]["pass"] for i in ids) |
| print(f"{label:<17}{t:>9} vs {b:<9} {100 * m:+6.2f} [{100 * lo:+6.2f}, {100 * hi:+6.2f}] flips {pf}/{fp} n={len(ids)}") |
|
|
| print("\nPer task (1 = pass, 0 = fail, T = agent timeout, E = other error, - = missing)") |
| tasks = sorted(set().union(*R.values())) |
| print(f"{'task':<40}" + "".join(f"{c:>9}" for c in CFGS)) |
| for t in tasks: |
| cells = [] |
| for c in CFGS: |
| x = R[c].get(t) |
| s = "-" if x is None else str(x["pass"]) + ({"AgentTimeoutError": "T", None: ""}.get(x["exc"], "E")) |
| cells.append(f"{s:>9}") |
| print(f"{t:<40}" + "".join(cells)) |
|
|
