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.