StrataSynth/stratasynth-agent-stress-test

StrataSynth Agent-Stress-Test

High-stakes conversational scenarios designed to stress-test dialogue agents: jealousy escalation, difficult performance reviews, and family estrangement. Includes gaslighting, accusation, deflection, and boundary-setting turns with full psychological metadata.

Dataset Details

Generated with StrataSynth — a platform for synthetic conversational data with psychological depth.

What makes this dataset different

• Explicit belief tracking: each turn records 12 psychological belief dimensions (trust, self-worth, hostility, etc.) with value + confidence
• Relationship trajectory: 4D dynamics (trust, tension, connection, dominance_balance) that evolve turn-by-turn
• Decision engine: intent, goal, and communication_act are computed before the LLM renders the text — cognition is separated from language
• Ground truth labels: affinity, communication style, belief consistency — computed without LLM (numpy/sklearn)
• Controlled noise: each dataset includes realistic noise turns (exaggeration, social lies, retractions) at 12% rate

Scenarios included

Job	Scenario	Description
job_71b1680f2379	ROM-02	Romantic jealousy escalation
job_79536e8a96c5	PRO-02	Workplace performance review
job_88d91faa2129	FAM-03	Family estrangement attempt

Schema

Each row is a conversation with the following fields:

{
  "conversation_id": "conv_xxx",
  "scenario_id": "FAM-01",
  "arc_type": "escalation | conflict | reconnection | ...",
  "language": "en",
  "num_turns": 20,
  "turns": [
    {
      "turn_id": "t001",
      "speaker": "persona_a | persona_b",
      "text": "...",
      "communication_act": "accusation | disclosure | validation | ...",
      "intent": "confront | deflect | reveal | ...",
      "goal": "protect_self | repair_bond | assert_control | ...",
      "emotional_state": {"tension": 0.7, "connection": 0.3, "vulnerability": 0.5},
      "relationship_state": {"trust": 0.4, "tension": 0.8, "connection": 0.3, "dominance_balance": 0.1},
      "belief_state": {"trust_other": 0.3, "self_worth": 0.6, ...},
      "belief_delta": {"trust_other": -0.05, ...},
      "noise_flags": []
    }
  ],
  "ground_truth": {...}
}

Usage

from datasets import load_dataset

ds = load_dataset("StrataSynth/stratasynth-agent-stress-test")
conv = ds["train"][0]
print(conv["scenario_id"], "—", conv["num_turns"], "turns")

for turn in conv["turns"]:
    print(f"[{turn['speaker']}] {turn['text']}")
    print(f"  act={turn['communication_act']} | intent={turn['intent']}")

License

CC BY 4.0 — free to use for research and commercial purposes with attribution.

Citation

@dataset{stratasynth_agent_stress_test,
  title  = {StrataSynth Agent-Stress-Test},
  author = {StrataSynth},
  year   = {2026},
  url    = {https://huggingface.co/datasets/StrataSynth/stratasynth-agent-stress-test}
}

Generated with StrataSynth v1.1.0 · stratasynth.com