wehub-resource-sync
a789495a98
FreeBSD Smoke / FreeBSD Smoke (x86_64) (push) Has been cancelled
CI / Quality Guardrails (push) Has been cancelled
CI / Build & Test (macos-latest) (push) Has been cancelled
CI / Build & Test (ubuntu-latest) (push) Has been cancelled
CI / Build & Test (windows-latest) (push) Has been cancelled
CI / Format (push) Has been cancelled
CI / PowerShell Syntax (push) Has been cancelled
CI / Windows Cross-Target Check (Linux) (push) Has been cancelled
108 lines
3.7 KiB
Python
108 lines
3.7 KiB
Python
#!/usr/bin/env python3
|
|
"""Aggregate a jcode Terminal-Bench Harbor job dir and compare against the
|
|
Claude Code + Opus 4.8 baseline.
|
|
|
|
Usage:
|
|
python scripts/tb_compare.py <jobs-dir> [baseline.tsv]
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import sys
|
|
from pathlib import Path
|
|
|
|
|
|
def load_baseline(path: Path) -> dict[str, float]:
|
|
out: dict[str, float] = {}
|
|
for line in path.read_text().splitlines():
|
|
line = line.strip()
|
|
if not line or line.startswith("#"):
|
|
continue
|
|
parts = line.split("\t")
|
|
if len(parts) < 4:
|
|
continue
|
|
task, _trials, _resolved, rate = parts[:4]
|
|
out[task] = float(rate)
|
|
return out
|
|
|
|
|
|
def collect_results(jobs_dir: Path) -> dict[str, list[float]]:
|
|
"""Map task_name -> list of per-trial rewards across all result.json files."""
|
|
results: dict[str, list[float]] = {}
|
|
for result_json in jobs_dir.rglob("result.json"):
|
|
try:
|
|
data = json.loads(result_json.read_text())
|
|
except Exception:
|
|
continue
|
|
stats = data.get("stats") or {}
|
|
evals = stats.get("evals") or {}
|
|
for _eval_name, ev in evals.items():
|
|
reward_stats = (ev.get("reward_stats") or {}).get("reward") or {}
|
|
for reward_str, trial_ids in reward_stats.items():
|
|
try:
|
|
reward = float(reward_str)
|
|
except ValueError:
|
|
continue
|
|
for trial_id in trial_ids:
|
|
# trial_id like "regex-log__abc123"
|
|
task = trial_id.split("__", 1)[0]
|
|
results.setdefault(task, []).append(reward)
|
|
return results
|
|
|
|
|
|
def main() -> int:
|
|
if len(sys.argv) < 2:
|
|
print(__doc__)
|
|
return 2
|
|
jobs_dir = Path(sys.argv[1]).expanduser()
|
|
baseline_path = Path(sys.argv[2]).expanduser() if len(sys.argv) > 2 else (
|
|
Path(__file__).parent / "tb_baseline_cc_opus48.tsv"
|
|
)
|
|
baseline = load_baseline(baseline_path)
|
|
results = collect_results(jobs_dir)
|
|
|
|
rows = []
|
|
jcode_resolved = jcode_trials = 0
|
|
regressions = []
|
|
for task in sorted(set(baseline) | set(results)):
|
|
rewards = results.get(task, [])
|
|
n = len(rewards)
|
|
passed = sum(1 for r in rewards if r >= 1.0)
|
|
rate = (100.0 * passed / n) if n else None
|
|
base = baseline.get(task)
|
|
jcode_resolved += passed
|
|
jcode_trials += n
|
|
flag = ""
|
|
if rate is not None and base is not None:
|
|
if rate < base:
|
|
flag = "REGRESSION"
|
|
regressions.append((task, base, rate))
|
|
elif rate > base:
|
|
flag = "gain"
|
|
rows.append((task, base, rate, passed, n, flag))
|
|
|
|
print(f"{'task':38} {'base%':>7} {'jcode%':>7} {'pass':>6} {'flag'}")
|
|
print("-" * 75)
|
|
for task, base, rate, passed, n, flag in rows:
|
|
base_s = f"{base:.1f}" if base is not None else "-"
|
|
rate_s = f"{rate:.1f}" if rate is not None else "n/a"
|
|
pass_s = f"{passed}/{n}" if n else "-"
|
|
print(f"{task:38} {base_s:>7} {rate_s:>7} {pass_s:>6} {flag}")
|
|
|
|
print("-" * 75)
|
|
if jcode_trials:
|
|
print(f"jcode micro-avg: {jcode_resolved}/{jcode_trials} = {100*jcode_resolved/jcode_trials:.1f}%")
|
|
base_resolved = sum(baseline.values())
|
|
print(f"baseline macro-avg: {base_resolved/len(baseline):.1f}% (CC+Opus4.8)")
|
|
if regressions:
|
|
print(f"\n{len(regressions)} REGRESSIONS vs baseline:")
|
|
for task, base, rate in regressions:
|
|
print(f" {task}: {base:.1f}% -> {rate:.1f}%")
|
|
else:
|
|
print("\nNo per-task regressions detected (for tasks with results).")
|
|
return 0
|
|
|
|
|
|
if __name__ == "__main__":
|
|
raise SystemExit(main())
|