five-desks / source /five_desks /environment.py
rca-umb's picture
Five Desks v1: tasks, calibration, source
1138ff6 verified
Raw History Blame Contribute Delete
4.03 kB
"""The Five Desks OpenEnv environment."""
import secrets
from typing import Any, Optional
from uuid import uuid4
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import State
from .calc import CalcError, evaluate, format_number
from .grading import answer_template, grade
from .models import DeskAction, DeskObservation
from .tasks import TASK_IDS, Instance, build
MAX_OUTPUT_CHARS = 6000
TOOLS_HELP = (
'Tools (reply with one JSON object per turn): {"tool": "list"}; {"tool": "read", "path": "<file>"}; '
'{"tool": "calc", "expr": "<expression>"} (numbers, + - * / // % **, parentheses, '
"sqrt ln log10 log2 exp abs round min max pow(a,b,m) floor ceil gcd lcm comb); "
'{"tool": "submit", "answer": {...}}.'
)
class FiveDesksEnvironment(Environment):
SUPPORTS_CONCURRENT_SESSIONS = True
def __init__(self):
self._state = State(episode_id=str(uuid4()), step_count=0)
self._instance: Instance | None = None
self._steps_left = 0
self._done = True
def reset(
self,
seed: Optional[int] = None,
episode_id: Optional[str] = None,
task_id: Optional[str] = None,
**kwargs: Any,
) -> DeskObservation:
task_id = task_id or TASK_IDS[0]
# Unseeded resets draw a fresh instance, so parallel rollouts see different data.
seed = seed if seed is not None else secrets.randbits(48)
self._instance = build(task_id, seed)
self._state = State(episode_id=episode_id or str(uuid4()), step_count=0)
self._steps_left = self._instance.max_steps
self._done = False
inst = self._instance
text = (
f"{inst.desk.upper()} DESK. {inst.question}\n"
f"Files: {', '.join(inst.files)}.\n"
f"Submit exactly: {{\"tool\": \"submit\", \"answer\": {answer_template(inst.parts)}}}\n"
f"{TOOLS_HELP} You have {inst.max_steps} actions; submit ends the episode."
)
return DeskObservation(output=text, steps_left=self._steps_left, done=False, reward=0.0)
def step(self, action: DeskAction, timeout_s: Optional[float] = None, **kwargs: Any) -> DeskObservation: # type: ignore[override]
if self._instance is None or self._done:
return DeskObservation(output="Episode is over. Call reset.", steps_left=0, done=True, reward=0.0)
self._state.step_count += 1
self._steps_left -= 1
inst = self._instance
if action.tool == "submit":
reward, scores = grade(inst.parts, action.answer)
self._done = True
correct = sum(1 for s in scores if s == 1.0)
return DeskObservation(
output=f"Submitted. {correct} of {len(scores)} answer parts correct.",
steps_left=self._steps_left, done=True, reward=round(reward, 4),
)
if action.tool == "list":
output = "\n".join(f"{name} ({len(body.splitlines())} lines)" for name, body in inst.files.items())
elif action.tool == "read":
name = action.path.strip().lstrip("./")
body = inst.files.get(name)
if body is None:
output = f"No file named {action.path!r}. Files: {', '.join(inst.files)}."
else:
output = body if len(body) <= MAX_OUTPUT_CHARS else body[:MAX_OUTPUT_CHARS] + "\n[truncated]"
else: # calc
try:
output = f"= {format_number(evaluate(action.expr))}"
except CalcError as exc:
output = f"calc error: {exc}"
if self._steps_left <= 0:
self._done = True
return DeskObservation(output=output + "\nOut of actions without a submission. Reward 0.",
steps_left=0, done=True, reward=0.0)
return DeskObservation(output=output, steps_left=self._steps_left, done=False, reward=0.0)
@property
def state(self) -> State:
return self._state