exp_rpt_bugsinpy-v2

A repaired version of DCAgent/exp_rpt_bugsinpy: 500 single-file Python bug-fixing tasks (BugsInPy-derived) with a robust, default-fail verifier that fixes the original's "reward is always 0" problem.

Why v2 exists

In the original dataset the per-task tests/test_solution.py files were LLM-generated and frequently malformed — undefined inspect/os imports, markdown fences, @pytest.main() decorators, bad bytes literals, references to APIs the buggy solution.py doesn't expose. Because tests/ is not agent-editable, any collection error in a test file makes the reward permanently 0 for that task, no matter what the agent does. A second class of tasks passed all tests on the buggy code, rewarding doing nothing.

This v2 makes every task produce a sensible reward signal.

Format

A single tasks.parquet with one row per task, columns path (bugsinpy-NNNN) and task_binary (gzipped tar of the task dir). Each task:

instruction.md            # the bug description (agent-facing)
metadata.json             # source / project / buggy+fixed commits
task.toml
environment/
  Dockerfile              # python:3.10-slim + pytest; bakes buggy solution + a baseline copy
  solution.py             # the BUGGY starter code (agent edits /app/solution.py)
tests/
  test.sh                 # robust default-fail verifier (see below)
  test_solution.py        # the per-task unit tests (mechanically repaired)
  test_state.py           # Harbor assertion: reward.txt == "1"; reports pass_ratio
  config.json             # test weights

Reward logic (tests/test.sh)

The verifier is default-fail and rewards 1 only when the agent produces a real, importable fix:

1. echo 0 > reward.txt (default-fail), set PYTHONPATH=/app.
2. No-op guard — /app/solution.py must differ from the baked-in /app/.buggy_solution.py. An unedited solution always scores 0 (kills the "buggy already passes all tests" leak).
3. Compile guard — python3 -m py_compile /app/solution.py must succeed.
4. Import guard — python3 -c "import solution" must succeed. Many buggy starters fail to import at module load (e.g. a missing import os); making the module importable is part of the task, so a still-broken import scores 0.
5. Run pytest /tests/test_solution.py and parse the summary:
   • Tests collect & run → reward 1 iff passed > 0 and failed == 0 (i.e. the agent's fix makes the whole suite pass); 0 otherwise.
   • Test file cannot be collected (irreparably broken independent of the solution) → fall back to the basic check: reward 1 iff changed + compiles + imports (the Option-B fallback).

pass_ratio = passed/(passed+failed) is written to /logs/verifier/pass_ratio.txt.

What was repaired

• Mechanically auto-repaired 13 test files (Option A): stripped embedded markdown fences/prose, removed the malformed @pytest.main() decorator, converted non-ASCII bytes literals to str, fixed from typing import Enum → from enum import Enum and from collections import MutableMapping → from collections.abc, and injected missing stdlib imports (os, sys, re, inspect, dataclasses.dataclass, abc.abstractmethod, typing.Optional, …) for names referenced-but-undefined.
• Generic fallback (Option B) applies to 6 tasks (~1%) whose test files reference APIs the solution does not expose (e.g. await outside async, tensorflow, missing solution attributes) and cannot be auto-repaired without a gold fix.

Reward-signal breakdown (against the buggy starter, 500 tasks)

Signal	Count	%
Test-driven (suite collects; fix → pass)	461	92%
Buggy solution fails to import (agent must repair import first)	31	6%
Generic fallback (broken test file → basic check)	6	1%
Timeout	2	<1%

Validation

Docker smoke test (build image per task, run tests/test.sh):

Scenario	Expected	Result
Unchanged buggy solution	0	0
Empty /app (no solution.py)	0	0
Changed but solution.py fails to import	0	0
Changed but tests still fail	0	0
Changed + all tests pass	1	1

Notes / limitations

• No gold solution/solve.sh is provided (parity with the source). The reward signal is verified structurally, not against a known-correct fix.
• The 6 generic-fallback tasks give reward 1 for any changed, importable solution (weaker signal); all other tasks use the real unit-test suite.
• Snapshot-safe: all 500 tasks share a single environment/Dockerfile.