tm23hgf/social-network-env

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

Hugging Face Space: https://huggingface.co/spaces/tm23hgf/social-network-env Core Env Logic: https://huggingface.co/spaces/tm23hgf/social-network-env/tree/main/socialnetworkenv

Github Link: https://github.com/TM23-sanji/socialnetwork Colab Link: https://colab.research.google.com/drive/1AXF0PdRhgbH63VP-AuziA9jL93z38Yh#scrollTo=74s-CR4kYpV9

InfoNet: Adversarial Social Network Detection Environment

1. Concept & Overview

InfoNet is a synthetic adversarial social network simulation where a detection agent learns to identify coordinated inauthentic behavior (propagandists) using only observable behavioral signals. The detector never sees ground-truth account roles, never sees content truth labels, and cannot rely on simple fact-checking.

The simulation runs 150 steps per episode. Every 5 steps, the detector reviews the network state and decides which accounts to flag. Propagandists adapt — changing topics and personas after being flagged — creating an adversarial co-training loop.

2. Agent Architecture

Agent Roster (10 Total)

Design Rationale: Clean Role Separation

• —Propagandists do not cite. If they did, they would look like amplifiers. The detector learns coordination patterns (who gets cited by whom), not role labels.
• —Neutrals do not cite. Clean separation means the detector learns timing and repetition patterns.
• —Amplifiers are the only citers. The detector doesn't know who they are citing until it sees the pattern.

3. Content Model

Every post has three attributes:

Key Design: True Facts + Framing

Propagandists use verifiably true facts with fear/outrage framing. The detector cannot win by keyword matching or truth detection — it must learn behavioral coordination.

4. Episode Mechanics

Episode Parameters

Simulation Loop

for step in 1..150:
    graph.step = step

    # Every agent acts (simultaneously)
    actions = _collect_actions()   # Neutral posts, Amplifier cites, Propagandist posts/reacts
    _execute_actions(actions)        # Add posts to graph, process amplifier assignments

    # Every 5 steps: detector reviews
    if step % 5 == 0:
        observation = build_observation()
        flagged = detector_llm(observation) OR fallback_policy()
        flagged = flagged[:2]        # Cap at 2 flags
        for acc in flagged:
            graph.flag_account(acc, step)
        reward_computer.compute_review_reward()

5. Agent Detail

NeutralAgent (5 accounts)

• —Posts from a neutral data pool at Poisson rates (0.30–0.45 per step depending on persona)
• —Personas: hothead, careful_journalist, clickbait, partisan, rando
• —Framing varies by persona (e.g., hothead → outrage/anger/fear dominant)
• —Posts from 301 topics, all topic-focused (no off-topic personal content)

AmplifierAgent (3 accounts)

• —Auto-recruited by propagandists when they post
• —Queued assignments: (prop_id, target_step, post_id)
• —At target_step, cites the propagandist's post
• —Falls back to citing neutral posts (80% chance) when no assignments mature
• —Not the primary detection target — amplifier citations help trace back to source

PropagandistAgent (2 accounts)

• —Has current_topic + angle_idx (0 or 1) + mimic_persona
• —Posts from the propaganda bank (301 topics × 2 angles × 5 personas)
• —50% posting probability per step when not flagged
• —When flagged (after 3-step delay), LLM decides action:
• —MIMIC(persona="hothead") — adopt new persona, flip angle
• —PIVOT(topic="Climate_Change") — switch topic, flip angle
• —BURN — quit immediately (goes silent forever)
• —WAIT — do nothing

6. Propaganda Bank

Location: data/episodes/propaganda_bank/

{Topic}/
  angle_0/{hothead|clickbait|partisan|rando|careful_journalist}.jsonl
  angle_1/{...}

• —301 topics × 2 angles × 5 personas × 30 samples = ~90,300 samples
• —Content: Neutral posts rewritten through angle-specific lens
• —Angle 0 "managed_stability": Institutions/experts handling it; critics are alarmist
• —Angle 1 "progressandstewardship": Markets/innovation solving problems; intervention causes harm

7. Detector Mechanics

Observation Format

At each review, the detector sees:

STEP 30 REVIEW OBSERVATION
============================================================

EPISODE SUMMARY:
- Steps completed: 30/150
- Total posts: 156
- Total citations: 77
- Total flagged accounts: 4
- Burned accounts: 0

SUSPICION TRACKER (all accounts):
  Account_1: rate=0.40, posts=13, cited=0, status=FLAGGED/ACTIVE
  Account_2: rate=1.00, posts=26, cited=26, status=FLAGGED/ACTIVE
  ...

RECENT ACTIVITY (last 8 steps = steps 23-30):
  Step 23: Account_1 POSTED [Cyber_Law] "Catch-22; stronger laws needed..."
  Step 23: Account_2 CITED (cites Account_10's post from step 19)
  ...

YOUR PREVIOUS FLAGS:
  - Step 15: Flagged Account_9
  - Step 20: Flagged Account_3
  ...

DECIDE: Flag up to 2 accounts, or WAIT.

Key insight: The detector sees the effects of PIVOT/MIMIC (topic changes, citation pattern shifts) but never sees the action JSON itself.

Output Schema

{
  "reasoning": "Account_1 and Account_9 consistently cite posts from Account_2, which was flagged in previous steps. This suggests coordinated amplification...",
  "flags": [
    {"account_id": 1, "confidence": 0.85},
    {"account_id": 9, "confidence": 0.72}
  ]
}

Parsing: If JSON parsing fails, fallback to empty flags (WAIT). The episode continues.

8. Reward Structure

Single training phase (no curriculum). All episodes use the same rewards.

Per-Review Reward

Active Propagandan Definition

An active propagandistan is one that:

1. 1.Posted within the last 5 steps, AND
2. 2.Was not flagged in the current review

F1 Metrics

Only propagandists are the positive class (amplifiers are FPs):

Precision = TP_prop / (TP_prop + FP_all)
Recall = TP_prop / (TP_prop + FN_prop)
F1 = 2 × Precision × Recall / (Precision + Recall)

9. Episode Output Files

ep_{NNNN}/
├── metadata.json              # Ground truth, reward summary, F1 metrics, final flags
├── final_state.json           # Full graph state including pivot log
├── detector_traces.jsonl     # SFT data: observation → {reasoning, flags}
└── propagandist_traces.jsonl  # Prop traces: prompt → {action, params, reasoning}

10. SFT Training Data

[image]

Only the detector is trained via SFT.

Format

System: You are a social network analyst tasked with detecting coordinated inauthentic behavior...

User (Observation):
<full observation text>

Assistant (Expected Output):
{"reasoning": "...", "flags": [{"account_id": 1, "confidence": 0.85}, ...]}

Current Data Quality

• —150 detector traces (5 episodes × 30 reviews each)
• —100% have valid JSON with proper structure
• —100% have chain-of-thought reasoning (avg 50+ words, min 27 words)
• —Structure: {"reasoning": str, "flags": [{"account_id": int, "confidence": float}]}

Planned Scale

• —Run 600 episodes with 3B model
• —Filter later based on quality metrics
• —All traces must have valid JSON with reasoning and flags fields

11. Key Constants

12. File Structure

infonet/
├── data/
│   ├── neutral_data_pool/
│   │   ├── config.json           # 301 topics
│   │   └── generated/            # 5 personas × 301 topics JSONL files
│   ├── episodes/
│   │   ├── src/                  # Simulation engine
│   │   │   ├── agents.py         # NeutralAgent, AmplifierAgent, PropagandistAgent
│   │   │   ├── config.py         # All constants
│   │   │   ├── episode.py        # EpisodeRunner — main loop
│   │   │   ├── graph.py          # GraphStateManager — central state
│   │   │   ├── observation.py    # ObservationBuilder — detector input
│   │   │   ├── prompts.py        # System prompts
│   │   │   ├── rewards.py        # RewardComputer
│   │   │   ├── schemas.py        # Pydantic models
│   │   │   ├── vllm_client.py   # LLM client with retry
│   │   │   └── detector.py       # Fallback policies
│   │   ├── episode_pool/         # Generated episodes
│   │   │   └── ep_{NNNN}/
│   │   ├── propaganda_bank/       # 301 topics × 2 angles × 5 personas
│   │   └── generate_episodes.py  # CLI entry point
├── README.md
└── requirements.txt

13. What Was Removed / Changed

14. How Agents Interact

Propagandistan posts
    ↓
Propagandan auto-recruits 2 amplifiers
    ↓
Amplifiers queued with delay
    ↓
Amplifiers cite at target step
    ↓
Detector reviews every 5 steps
    ↓
Detector flags suspicious accounts
    ↓
Propagandan sees flag after 3-step delay
    ↓
Propagandan LLM reacts: MIMIC/PIVOT/BURN/WAIT
    ↓
Topic/persona changes → new coordination patterns
    ↓
Detector observes behavioral changes

15. Current Status

• —Episode generation: Working (3B model via vLLM)
• —Reward structure: Finalized
• —SFT data: 150 traces from 5 episodes
• —Scale-up: Ready for 600 episodes
• —Training pipeline: Not yet implemented

Document version: 2.0 Last updated: After implementation consolidation