| """One-step numerical-answer environment for the combined STEM dataset.""" |
| from __future__ import annotations |
|
|
| import json |
| from pathlib import Path |
| import random |
| import tempfile |
| from typing import Any |
|
|
| try: |
| from .grade import DOMAINS, TASKS, grade_task, package_path |
| except ImportError: |
| from grade import DOMAINS, TASKS, grade_task, package_path |
|
|
|
|
| class STEMEnv: |
| """Select a fixed task, a domain, or any of the 40 tasks. Not a security sandbox.""" |
|
|
| def __init__(self, task_id: str | None = None, domain: str | None = None): |
| if domain is not None: |
| domain = domain.upper() |
| if domain not in DOMAINS: |
| raise ValueError("Unknown domain: " + domain) |
| if task_id is not None and task_id not in TASKS: |
| raise ValueError("Unknown task ID: " + task_id) |
| if task_id is not None and domain is not None and TASKS[task_id]["domain"] != domain: |
| raise ValueError("Task does not belong to the requested domain") |
| self.task_ids = [tid for tid, info in TASKS.items() |
| if (domain is None or info["domain"] == domain) |
| and (task_id is None or tid == task_id)] |
| self.task_id: str | None = None |
| self.terminated = False |
|
|
| def reset(self, seed: int | None = None) -> tuple[dict[str, Any], dict[str, Any]]: |
| self.task_id = random.Random(seed).choice(self.task_ids) |
| self.terminated = False |
| task = TASKS[self.task_id] |
| base = package_path(task["path"]) |
| observation = { |
| "task_id": self.task_id, |
| "domain": task["domain"], |
| "instruction": (base / "instruction.md").read_text(encoding="utf-8"), |
| "instance": json.loads((base / "instance.json").read_text(encoding="utf-8")), |
| } |
| return observation, {"task_id": self.task_id, "domain": task["domain"], "max_steps": 1} |
|
|
| def step(self, action: dict[str, Any] | str | bytes) -> tuple[dict[str, Any], float, bool, bool, dict[str, Any]]: |
| if self.task_id is None or self.terminated: |
| raise RuntimeError("Call reset() before each one-action episode") |
| if isinstance(action, dict): |
| raw = json.dumps(action, ensure_ascii=False, allow_nan=False).encode("utf-8") |
| elif isinstance(action, str): |
| raw = action.encode("utf-8") |
| elif isinstance(action, bytes): |
| raw = action |
| else: |
| raise TypeError("Action must be a JSON object, string, or bytes") |
| if len(raw) > TASKS[self.task_id]["candidate_maximum_bytes"]: |
| raise ValueError("Candidate exceeds size limit") |
| with tempfile.TemporaryDirectory(prefix="stem-answer-") as scratch: |
| candidate = Path(scratch) / "candidate.json" |
| candidate.write_bytes(raw) |
| result = grade_task(self.task_id, candidate) |
| self.terminated = True |
| return {"task_id": self.task_id, "passed": result["passed"]}, result["reward"], True, False, result |
|
|
