---
license: mit
language:
- en
task_categories:
- text-generation
tags:
- rl-environment
- verifiers
- code
- bug-fixing
- swe-bench
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
---

# Polyglot HardBug RL

A Hugging Face RL environment dataset of 10 **expert-level, multi-language
bug-fixing tasks** for training and evaluating coding agents. Each task is a
self-contained SWE-bench-style instance: a real buggy project, a precise problem
statement, hidden verification tests, and a gold patch.

Every task was authored to reproduce a *class* of subtle defect that standard
datasets miss - the bugs are semantically rich (concurrency, three-valued SQL
logic, numeric type systems, Unicode canonicalization) rather than syntactic
slips, and each one was verified bidirectionally: the buggy repository fails
exactly the declared `fail_to_pass` tests, and the gold patch makes the whole
suite pass.

## Tasks by language

| Language | Tasks |
| --- | --- |
| javascript | 1 |
| python | 7 |
| sql | 1 |
| typescript | 1 |

## Row schema

| Column | Description |
| --- | --- |
| `task_id` | stable task identifier (also the HF row id) |
| `title` | one-line description of the defect |
| `language` | python / javascript / typescript / sql |
| `category` | defect class, e.g. concurrency-async, sql-semantics |
| `difficulty` | difficulty rating (all rows: expert) |
| `source` | provenance (original-synthetic: authored for this dataset) |
| `license` | per-row license (MIT) |
| `created` | authoring date |
| `prompt` | the agent-facing problem statement (markdown) |
| `repo_files` | the buggy workspace: list of `{filename, content}` entries |
| `test_files` | the hidden verification suite (same entry format) |
| `gold_patch` | unified diff (`a/`/`b/` prefixes) fixing the bug |
| `test_command` | argv that runs the hidden suite inside the sandbox |
| `fail_to_pass` | tests that fail on the buggy repo and pass after the fix |
| `pass_to_pass` | tests that must keep passing (regression guard) |
| `time_limit_s` | test-run budget in seconds |
| `sandbox_image` | runtime image for the task (python or node toolchain) |
| `sandbox_system_deps` | extra system packages the image needs (e.g. tzdata) |

- `repo_files` entries are sorted by filename; rebuild a dict with
  `{e["filename"]: e["content"] for e in repo_files}`. The agent edits this
  state (typically by emitting a unified diff).
- `test_files` holds the hidden verification suite, injected by the harness
  after the agent finishes, so any test edits the agent made are overwritten.
- `gold_patch` is `patch -p1` compatible and transforms `repo_files` into the
  fixed state.

## Verification protocol

The suite is TAP-based. Each test prints `ok N - name` or `not ok N - name`
with a diagnostic; the process exits non-zero if any test failed. A solution is
**resolved** iff every `fail_to_pass` test passes and every `pass_to_pass` test
still passes.

## Intended use

Load with the companion verifiers environment package
(`polyglot-hardbug-rl/environment/`), or drive the rows with your own harness:

```python
from datasets import load_dataset

rows = load_dataset("mondk/polyglot-hardbug-rl", split="train")
for row in rows:
    write_workspace(row["repo_files"])          # agent edits this
    apply_patch(agent_patch)                    # model's diff
    inject(row["test_files"])                   # hidden tests
    stdout = run(row["test_command"], timeout=row["time_limit_s"])
    score = score_tap(stdout, row["fail_to_pass"], row["pass_to_pass"])
```

## Repository layout

| path | contents |
| --- | --- |
| `data/` | the published dataset (`train.jsonl`, `train-00000-of-00001.parquet`) |
| `tasks/` | one directory per task: `problem.md`, `repo/` (buggy), `fixed/` (gold), `tests/` (hidden TAP suite), `meta.json`, `run.json` |
| `scripts/build_dataset.py` | rebuilds `data/` and this card from `tasks/`; enforces the schema and the pure-ASCII English gate |
| `scripts/validate_dataset.py` | end-to-end check: buggy fails exactly `fail_to_pass`, gold patch makes the whole suite pass; writes `validation_report.md` |
| `environment/` | the verifiers v1 environment package (`polyglot_hardbug`) plus a standalone local runner |
| `validation_report.md` | result of the latest `validate_dataset.py` run |

## License

MIT (see `LICENSE`). Task content is original synthetic work created for this
dataset; no third-party source code is included.