qed-math-openenv/openenv.yaml at main · 3xcaffeine/qed-math-openenv · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
spec_version: 1
name: qed_math_env
type: space
runtime: fastapi
app: server.app:app
port: 8000

description: >
  Mathematical proof generation and evaluation environment.
  Agents receive math problems, submit proofs, and receive LLM-based
  rubric grading (0-7 scale) with normalized rewards.
  Answer-mode uses a process-based verifier service for concurrent rollout
  safety (timeouts, retries, backpressure, worker restart).
version: "0.1.0"

# Environment configuration defaults (overridable via QEDMathConfig)
environment:
  dataset_path:
    hub_id: lm-provers/FineProofs-RL-test
    split: test[:10]
  grader_model: gemini-3-pro
  prompt_name: v2
  custom_reward_threshold: false
  discount_factor: 1.0
  buffer_tokens: 0
  max_tokens: 0
  verifier_workers: 4
  verifier_queue_size: 128
  verifier_request_timeout_seconds: 5.0
  verifier_max_retries: 1
  verifier_strict: true
  verifier_numeric_precision: 5
  verifier_float_rounding: 10

# Rubric definition
rubric:
  type: llm_judge
  name: MathProofRubric
  scale: "0-7"
  normalization: "score / 7.0"
  thresholding: "optional (collapses 1-5 -> 1)"

# MCP tools exposed by this environment
tools:
  - name: get_problem
    description: Return current problem statement and metadata
  - name: submit_proof
    description: Submit a proof for LLM-based rubric grading
    parameters:
      - proof (str, required)
  - name: get_grading_guidelines
    description: Return the rubric/marking scheme for the current problem
  - name: list_task_ids
    description: Return ordered task ids loaded by the current dataset configuration

# Verifier metrics emitted per grading call
metrics:
  verifier:
    - verifier/rollouts/success
    - verifier/rollouts/failure
    - verifier/failures/timeout
    - verifier/failures/rate_limit
    - verifier/failures/no_input
    - verifier/failures/no_score_tag
    - verifier/failures/all_attempts_failed
    - verifier/failures/num_retries
    - verifier/runtime/latency_per_request
    - verifier/requests/count
    - verifier/requests/latency_ms
    - verifier/requests/timeout_count
    - verifier/requests/error_count
    - verifier/queue/depth
    - verifier/cache/hit_rate
    - verifier/workers/restart_count
    - verifier/workers/worker_restarted
    - verifier/workers/heartbeat_lag_ms
    - verifier/runtime/input_tokens
    - verifier/runtime/output_tokens
  reward:
    - reward/base
    - reward/shaped
    - reward/score_raw
    - reward/overlong_penalty
  episode:
    - episode/attempt_number
    - episode/is_correct
    - episode/problem_type
    - episode/dataset_source