{"data":{"kind":"file","path":"README.md","version_id":"y765bnfx4f9126uqvs9vlgcz","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":5939,"modified_at":"2026-08-07T06:29:56.646000","content_hash":"be13984fae24c5980b14de943246ed05ac6a6ae9bc3e8515e45e6b7e53eaa4ca"},"entries":[],"content":"# numpy-scipy-env\r\n\r\n### Overview\r\n- **Environment ID**: `numpy-scipy-env`\r\n- **Short description**: RL environment for numpy/scipy numerical tasks using expected_output comparison\r\n- **Tags**: numpy, scipy, linear-algebra, scientific-computing\r\n\r\n### Datasets\r\n- **Primary dataset(s)**: `eltociear/numpy-scipy-tasks-v1`\r\n- **Source links**: [HuggingFace Dataset](https://huggingface.co/datasets/eltociear/numpy-scipy-tasks-v1)\r\n- **Split sizes**: train (40 examples)\r\n\r\n### Task Categories\r\n\r\n| Category | Description |\r\n|----------|-------------|\r\n| array_ops | Reshape, sort, cumsum, clip, masking, argsort, transpose, axis reductions, standardisation, diff |\r\n| linalg | matmul, inverse, symmetric eigenvalues, Cholesky, solve, SVD, QR, matrix exponential, pinv, determinant |\r\n| statistics | mean/median/std, percentiles, z-scores, correlation, covariance, chi-square, linear regression, ranks |\r\n| signal | real FFT magnitude and frequencies, convolution, DCT-II, linear detrend, peak finding |\r\n| optimization | polyfit, lstsq, BFGS minimisation, Brent root finding |\r\n\r\n### Task\r\n- **Type**: Multi-turn tool use (default `max_turns=5`)\r\n- **Rubric overview**: Binary pass/fail using `numpy.testing.assert_allclose` (rtol 1e-6, atol 1e-8) against the reference array\r\n\r\n### Quickstart\r\n\r\n```bash\r\nuv run vf-eval numpy-scipy-env -p prime -m openai/gpt-5.4-nano -s\r\n```\r\n\r\nFull sweep:\r\n\r\n```bash\r\nuv run vf-eval numpy-scipy-env -p prime -m openai/gpt-5.4-mini -n 40 -r 3 -s\r\n```\r\n\r\n### Environment Arguments\r\n\r\n| Arg | Type | Default | Description |\r\n|-----|------|---------|-------------|\r\n| `split` | str | `\"train\"` | Dataset split to use |\r\n| `dataset_name` | str | `\"eltociear/numpy-scipy-tasks-v1\"` | HuggingFace dataset name |\r\n| `max_turns` | int | `5` | Maximum interaction turns per task |\r\n\r\n### Tools Available\r\n- `execute_code(code: str)` — run Python in the sandbox; the task's input arrays are preloaded by name and `result` persists across turns\r\n- `bash(command: str)` — run shell commands in the sandbox\r\n\r\n### Grading\r\n\r\nThe reference answer never enters the sandbox — it is held host-side and compared after the\r\nrollout, so the model cannot read the answer key.\r\n\r\nComparison is `assert_allclose` with a tolerance, **not** exact equality (the sibling\r\n`pandas_env` uses exact equality). The answers here come from eigensolvers, matrix\r\nexponentials, FFTs and optimisers, whose last bits legitimately differ across BLAS builds,\r\nCPU architectures and library versions; grading those bit-exactly would fail correct solutions\r\nfor reasons that have nothing to do with the model.\r\n\r\nThe tolerance is prevented from laundering genuinely wrong answers: **shape is compared\r\nexactly**, and an integer-valued reference **requires** an integer answer.\r\n\r\nA wrong answer scores `0.0` silently. A broken scorer scores `0.0` and additionally sets\r\n`state[\"scoring_error\"]`, so eval runs can tell \"the model was wrong\" from \"the harness\r\nfailed\".\r\n\r\n### Building the dataset\r\n\r\n```bash\r\npython build_tasks.py --verify          # check every task\r\npython build_tasks.py --out train.jsonl # regenerate\r\n```\r\n\r\nTasks are (deterministic input arrays, instruction, reference solution); expected outputs are\r\n**computed** from the reference, never hand-written. `--verify` independently checks each task\r\nfor clean execution, real-valued output, determinism across two runs, finiteness (no NaN/inf in\r\nthe answer key), non-emptiness, non-identity, and an exact serialisation round-trip including\r\ndtype and shape. All 40 pass.\r\n\r\n## Environment arguments\r\n\r\n`load_environment()` deliberately exposes very little: the program's guidance is that an\r\nenvironment should have one correct way to be run, so nothing about the prompts, the parsing or\r\nthe grading is configurable.\r\n\r\n| Argument | Default | Meaning |\r\n|---|---|---|\r\n| `split` | `'train'` | Dataset split to load. |\r\n| `dataset_name` | `'eltociear/numpy-scipy-tasks-v1'` | Hugging Face dataset of tasks. Change only to point at a fork. |\r\n| `max_turns` | `5` | Tool-use turns the model gets before the rollout ends. |\r\n| `**kwargs` | — | Passed through to the underlying `SandboxEnv`. |\r\n\r\n## Reward rubric\r\n\r\n| Reward function | Weight | What it returns |\r\n|---|---|---|\r\n| `correctness` | 1.0 | Binary: 1.0 when the answer matches the reference, else 0.0. |\r\n\r\nThere is no LLM judge and no partial credit. The score is computed on the host in\r\n`post_rollout` and read back by the rubric, so the reward is a deterministic function of the\r\nvalues the model left in `result`. A **harness** failure (dead sandbox, unreadable read-back)\r\nalso scores 0.0 but additionally sets `state[\"scoring_error\"]`, so an eval run can tell a\r\nbroken harness from a wrong answer instead of blaming the model.\r\n\r\n## Dependencies\r\n\r\n`datasets>=4.1.0`, `numpy>=1.26.0`, `scipy>=1.11.0`, `verifiers>=0.1.8`\r\n\r\n## Sample `vf-eval` usage\r\n\r\n```bash\r\nuv run vf-install numpy-scipy-env\r\nuv run vf-eval -s numpy-scipy-env -m gpt-4.1 -n 5 -r 3\r\nuv run vf-tui                      # inspect the outputs/ folder it writes\r\n```\r\n\r\n### Known limitation\r\n\r\nThe Docker **sandbox transport** has not been executed — `docker run`, `pip install`, and the\r\nmodel's code running inside the container — because that needs a Docker runtime and an\r\ninference provider key, neither available where this was authored.\r\n\r\nEverything else is exercised. `environments/test_scoring_path.py` runs this environment's\r\n`post_rollout` and `Rubric` with the sandbox mocked out, asserting that a correct answer scores\r\n1.0, a well-formed wrong answer scores 0.0, and a dead sandbox scores 0.0 *and* sets\r\n`scoring_error` so a harness failure is never mistaken for a bad model. It also checks that\r\nwhat `build_tasks.py` emits is exactly what the scorer expects. The environment mirrors the\r\nstructure of `polars_env` (already accepted into the Environments Program) and imports cleanly\r\nagainst `verifiers` 0.2.1.\r\n","encoding":"utf-8","truncated":false,"total_bytes":5939},"status":null}