Spaces:
Sleeping
Sleeping
File size: 9,409 Bytes
e5e36db f1e8d43 5ff0156 f1e8d43 5ff0156 f1e8d43 5ff0156 e5e36db | 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 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 | # ============================================================================
# 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" }
|