| """How exactly does the untrained model fail the library bug fixes? Rebuild each calibration episode's planted bug from |
| its seed and check what the model's edits touched. |
| |
| Usage: .venv/bin/python scripts/wrong_fixes.py > reports/wrong-answers/bugfix.txt |
| """ |
|
|
| import collections |
| import glob |
| import json |
| import random |
| import sys |
| from pathlib import Path |
|
|
| ROOT = Path(__file__).resolve().parents[1] |
| sys.path.insert(0, str(ROOT)) |
| sys.path.insert(0, str(ROOT / "scripts")) |
|
|
| from calibrate import first_json_object |
| from codedesk.environment import build |
| from codedesk.interface import Interface |
|
|
| FORMAT_ENDS = {"json_plus_text", "no_json", "code_fence", "not_object", "empty", "truncated", "budget"} |
|
|
|
|
| def main(): |
| rows = [] |
| for f in sorted(glob.glob(str(ROOT / "reports/calibration-code/*/episodes.jsonl"))): |
| for line in open(f): |
| ep = json.loads(line) |
| if not ep["task_id"].startswith("bugfix-") or "seed" not in ep: |
| continue |
| task = build(ep["task_id"], ep["seed"]) |
| try: |
| tools = task.tools() |
| iface = Interface(tools, random.Random(f"names:{ep['task_id']}:{ep['seed']}")) |
| back_tool = {new: old for old, new in iface.tool_names.items()} |
| back_args = {} |
| for old, new in zip(tools, iface.tools): |
| back_args[old.name] = {n: o for o, n in zip(old.args, new.args)} |
| m = task.m |
| bad_line = task.original.splitlines()[m["line"] - 1].strip() |
| mutated_line = (ROOT / "x").name and mutated(task) |
| edits, ran_tests, read_target = [], 0, False |
| for t in ep["turns"]: |
| o = first_json_object(t["reply"]) or {} |
| tool = back_tool.get(str(o.get("tool", "")).strip()) |
| args = {back_args.get(tool, {}).get(k, k): v for k, v in (o.get("args") or {}).items()} |
| path = str(args.get("path", "")).lstrip("./") |
| if tool in ("edit_file", "write_file"): |
| text = str(args.get("old", "")) + "\n" + str(args.get("content", "")) |
| edits.append((path, mutated_line in text or bad_line in text, str(args.get("new", "")))) |
| if tool == "run_tests": |
| ran_tests += 1 |
| if tool == "read_file" and path == m["file"]: |
| read_target = True |
| row = {"task_id": ep["task_id"], "group": task.group, "lib": m["lib"], "level": task.level, |
| "reward": ep["reward"], "end": ep.get("end"), "turns": len(ep["turns"])} |
| lib_edits = [e for e in edits if not e[0].startswith(("tests", "test")) and "/test" not in e[0]] |
| test_edits = [e for e in edits if e not in lib_edits] |
| if ep["reward"] >= 0.999: |
| row["status"] = "solved" |
| elif ep.get("end") in FORMAT_ENDS: |
| row["status"] = "format ending" + (" after editing the bug's line" if any(e[1] for e in lib_edits) else "") |
| elif ep.get("end") == "schema_invalid": |
| row["status"] = "action rejected by the schema" |
| elif not lib_edits and test_edits: |
| row["status"] = "edited only tests" |
| elif not lib_edits: |
| row["status"] = "no edit to the library" + (" (never opened the buggy file)" if not read_target else |
| " (opened the buggy file)") |
| elif any(e[0] == m["file"] and e[1] for e in lib_edits): |
| row["status"] = "edited the bug's line, fix still wrong" |
| elif any(e[0] == m["file"] for e in lib_edits): |
| row["status"] = "edited the right file, wrong place" |
| else: |
| row["status"] = "edited a different file" |
| row["ran_tests"] = ran_tests |
| rows.append(row) |
| finally: |
| task.close() |
| report(rows) |
|
|
|
|
| def mutated(task): |
| from codedesk.families.bugfix import mutate |
| return mutate(task.original, task.m).splitlines()[task.m["line"] - 1].strip() |
|
|
|
|
| def report(rows): |
| print(f"bug-fix episodes rebuilt: {len(rows)}\n") |
| wrong = [r for r in rows if r["status"] != "solved"] |
| print("All failed episodes:") |
| for k, v in collections.Counter(r["status"] for r in wrong).most_common(): |
| print(f" {k:55s} {v:4d} {v/len(wrong):.0%}") |
| print("\nBy group (domain library):") |
| for g in sorted({r["group"] for r in rows}): |
| G = [r for r in rows if r["group"] == g] |
| W = [r for r in G if r["status"] != "solved"] |
| top = ", ".join(f"{k} {v}" for k, v in collections.Counter(r["status"] for r in W).most_common(3)) |
| print(f" {g:11s} solved {len(G)-len(W)}/{len(G)} {top}") |
| for lv in (1, 2): |
| L = [r for r in rows if r["level"] == lv] |
| if L: |
| print(f"\nlevel {lv}: solved {sum(r['status']=='solved' for r in L)}/{len(L)}; ran tests in " |
| f"{sum(r['ran_tests']>0 for r in L)/len(L):.0%} of episodes") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
