Spaces:
Sleeping
Sleeping
| # ============================================================================ | |
| # openenv.yaml v1.1.0 | |
| # OpenEnv environment specification — passes `openenv validate` | |
| # ============================================================================ | |
| name: information-integrity-env | |
| version: "1.1.0" | |
| author: openenv-content-moderation | |
| description: > | |
| A deterministic, benchmarkable content moderation environment where agents | |
| verify claims, reason over pre-built evidence sources, and make explainable | |
| moderation decisions (allow / flag / remove). | |
| Evaluation runs fully offline — no external API calls during episodes. | |
| Three difficulty tiers (easy / medium / hard) with static scenario datasets | |
| ensure reproducible, comparable benchmarks across agent runs. | |
| license: Apache-2.0 | |
| # --------------------------------------------------------------------------- | |
| # Server | |
| # --------------------------------------------------------------------------- | |
| server: | |
| type: http | |
| reset_endpoint: POST /reset | |
| step_endpoint: POST /step | |
| health_endpoint: GET /health | |
| tasks_endpoint: GET /tasks | |
| grade_endpoint: POST /grade | |
| collect_endpoint: POST /collect # dataset creation only; not used during eval | |
| state_endpoint: GET /state/{session_id} | |
| # --------------------------------------------------------------------------- | |
| # Observation space | |
| # --------------------------------------------------------------------------- | |
| observation: | |
| type: object | |
| description: > | |
| The agent observes the content under review, extracted claims (with live | |
| verification status), a pre-built evidence pool, and a running interaction | |
| history. Ground truth is never exposed. | |
| properties: | |
| content: | |
| type: string | |
| description: The piece of content under moderation review. | |
| claims: | |
| type: array | |
| description: Verifiable claims extracted from the content. | |
| items: | |
| type: object | |
| properties: | |
| id: { type: string } | |
| text: { type: string } | |
| verification_status: { type: string, enum: [unverified, true, false, partially_true, unknown] } | |
| confidence: { type: number, minimum: 0.0, maximum: 1.0 } | |
| verified_by_sources: { type: array, items: { type: string } } | |
| available_sources: | |
| type: array | |
| description: Pre-built evidence pool — fully deterministic, no live fetching. | |
| items: | |
| type: object | |
| properties: | |
| id: { type: string } | |
| title: { type: string } | |
| text: { type: string } | |
| url: { type: string } | |
| stance: { type: string, enum: [supporting, contradicting, neutral] } | |
| reliability: { type: number, minimum: 0.0, maximum: 1.0 } | |
| source_type: { type: string } | |
| interaction_history: | |
| type: array | |
| description: Log of all previous actions and their results this episode. | |
| status_flags: | |
| type: object | |
| properties: | |
| decision_made: { type: boolean } | |
| sufficient_evidence: { type: boolean } | |
| premature_decision: { type: boolean } | |
| redundant_actions: { type: integer } | |
| max_steps_reached: { type: boolean } | |
| step_count: { type: integer } | |
| episode_id: { type: string } | |
| task_difficulty: { type: string, enum: [easy, medium, hard] } | |
| # --------------------------------------------------------------------------- | |
| # Action space | |
| # --------------------------------------------------------------------------- | |
| actions: | |
| verify_claim: | |
| description: Investigate a specific claim and receive its verification status. | |
| required_fields: [claim_id] | |
| optional_fields: [reasoning] | |
| example: | |
| type: verify_claim | |
| claim_id: claim_001 | |
| reasoning: "This is the central assertion — verify first." | |
| check_source: | |
| description: Read a source and receive its full text, stance, and reliability. | |
| required_fields: [source_id] | |
| optional_fields: [reasoning] | |
| example: | |
| type: check_source | |
| source_id: src_003 | |
| reasoning: "WHO source — likely authoritative." | |
| decide: | |
| description: > | |
| Make the final moderation decision, ending the episode. | |
| Must verify >= 1 claim and check >= 2 sources first (or incur premature penalty). | |
| required_fields: [decision] | |
| optional_fields: [reasoning] | |
| example: | |
| type: decide | |
| decision: remove | |
| reasoning: "All claims false per three high-reliability sources." | |
| # --------------------------------------------------------------------------- | |
| # Reward | |
| # --------------------------------------------------------------------------- | |
| reward: | |
| type: float | |
| range: [-2.0, 2.0] | |
| description: > | |
| Dense step-level reward. Non-sparse by design — agents receive positive | |
| signals for every unique claim verified and source checked. | |
| See reward_config.yaml for configurable magnitudes. | |
| components: | |
| verify_claim_reward: "+0.20 per unique claim verified" | |
| check_source_reward: "+0.20 per unique source checked" | |
| multi_source_bonus: "+0.20 when >= 3 unique sources checked" | |
| decision_correct: "+1.00 exact match with ground truth" | |
| decision_partial: "+0.50 one step away (flag<->remove, flag<->allow)" | |
| decision_wrong: "0.00" | |
| premature_decision: "-0.50 deciding before sufficient evidence" | |
| redundant_action: "-0.10 re-verifying / re-checking same id" | |
| step_inefficiency: "-0.05 per step beyond efficient threshold" | |
| over_aggression: "-0.30 choosing remove when ground truth is allow/flag" | |
| # --------------------------------------------------------------------------- | |
| # Tasks (deterministic datasets — no live API calls) | |
| # --------------------------------------------------------------------------- | |
| tasks: | |
| - name: easy | |
| grader: tasks.easy.grader.grade_episode | |
| - name: medium | |
| grader: tasks.medium.grader.grade_episode | |
| - name: hard | |
| grader: tasks.hard.grader.grade_episode | |
| # --------------------------------------------------------------------------- | |
| # Grader | |
| # --------------------------------------------------------------------------- | |
| grader: | |
| output_range: [0.0, 1.0] | |
| dimensions: | |
| decision_score: | |
| description: Correctness of the final moderation call (partial credit for adjacent decisions) | |
| evidence_score: | |
| description: Fraction of available sources consulted | |
| reasoning_score: | |
| description: Fraction of claims fully verified before deciding | |
| efficiency_score: | |
| description: Steps used relative to difficulty-specific ideal maximum | |
| # --------------------------------------------------------------------------- | |
| # Evaluation protocol | |
| # --------------------------------------------------------------------------- | |
| evaluation: | |
| description: > | |
| Run inference.py in benchmark mode to evaluate an agent across all 15 scenarios. | |
| No external API calls are made during evaluation episodes. | |
| command: python inference.py --benchmark --out results.json | |
| expected_output_format: | | |
| difficulty=easy avg_score=X.XXXX avg_reward=X.XXXX n=5 | |
| difficulty=medium avg_score=X.XXXX avg_reward=X.XXXX n=5 | |
| difficulty=hard avg_score=X.XXXX avg_reward=X.XXXX n=5 | |
| tier_validation: easy_avg > medium_avg > hard_avg | |
| runtime_budget: "< 20 minutes on 2 vCPU / 8 GB RAM" | |
| logging_protocol: "[START] / [STEP] / [END]" | |
| # --------------------------------------------------------------------------- | |
| # Data pipeline (dataset creation only — not used during eval) | |
| # --------------------------------------------------------------------------- | |
| data_pipeline: | |
| description: > | |
| DataCollector is used ONLY to build the static scenario datasets in | |
| tasks/{difficulty}/scenarios.json. It is NOT called during evaluation episodes. | |
| backends: | |
| - { name: serpapi, env_key: SERPAPI_KEY } | |
| - { name: newsapi, env_key: NEWSAPI_KEY } | |
| - { name: brave_search, env_key: BRAVE_SEARCH_KEY } | |
| - { name: google_factcheck, env_key: GOOGLE_FACTCHECK_KEY } | |
| - { name: duckduckgo, env_key: null, note: "No key required; rate-limited fallback" } | |
| cache: | |
| backend: redis | |
| env_key: REDIS_URL | |
| default: redis://localhost:6379 | |
| ttl_seconds: 86400 | |
| # --------------------------------------------------------------------------- | |
| # Constraints | |
| # --------------------------------------------------------------------------- | |
| constraints: | |
| - Ground truth is never exposed in the Observation | |
| - No external API calls occur during env.step() or env.reset() | |
| - All reward computations are deterministic | |
| - Agents must gather evidence before deciding (premature decision penalty applies) | |
| - Redundant actions (re-checking same id) are penalised | |
| - Scenario datasets are static JSON — reproducible across runs | |
| # --------------------------------------------------------------------------- | |
| # Baseline | |
| # --------------------------------------------------------------------------- | |
| baseline: | |
| script: inference.py | |
| mode: --benchmark | |
| agent_type: ReAct (LLM-driven) | |
| logging_protocol: "[START] / [STEP] / [END]" | |
| model_env_var: MODEL_NAME | |
| api_base_env_var: API_BASE_URL | |
| token_env_var: HF_TOKEN | |
| expected_scores: | |
| easy: { avg_score: "~0.78", avg_reward: "~1.20" } | |
| medium: { avg_score: "~0.63", avg_reward: "~0.90" } | |
| hard: { avg_score: "~0.52", avg_reward: "~0.65" } | |