staqc-v2

# laion/staqc-v2 LLM-judge-verified version of `mlfoundations-dev/staqc-sandboxes`. Each task follows the `nemotron_gym` LLM-judge verifier contract: * `instruction.md` — the original task instruction, with submission guidance appended directing the agent to write its answer to `/app/response.txt`. * `tests/test_state.py` — a pytest-runnable LLM judge that reads `/app/response.txt` (legacy `/app/answer.txt` fallback), reads `tests/verifier_data.json` for the instruction + rubric, calls `litellm.completion` (default `openai/gpt-4o-mini`, temperature=0) with the rubric, parses `\boxed{<score>}`, and writes the float score to `/logs/verifier/reward.txt`. * `tests/test.sh` — defaults reward to 0, then runs `python3 -m pytest /tests/test_state.py`. * `tests/verifier_data.json` — the task instruction + a per-dataset rubric. * `environment/Dockerfile` — `ubuntu:24.04` + python3 + pip + openai + pytest + litellm. * `task.toml` — `LLM_JUDGE_TASK_TOML`, so `OPENAI_API_KEY` / `JUDGE_MODEL` propagate into the verifier container via `[verifier].env` (`OPENAI_API_KEY` is required at trial time). Stack Overflow question->code pairs (STAQC). The agent is asked to answer the programming question; graded by an LLM judge on correctness / completeness / clarity of the code answer. **Rubric:** correctness / completeness / clarity `10000` tasks.