cyberandy/sangue-e-grafi-agent-traces

🩸 Sangue e Grafi — Agent Traces Dataset

100 recorded agent traces showing a KG-grounded agent solving adversarial Italian inheritance-law scenarios.

Dataset Description

This dataset contains 100 agent trace recordings from the Sangue e Grafi project. Each trace captures a complete reasoning episode: a small (4B) language model navigating a kinship knowledge graph via tool calls to answer adversarial inheritance-law questions in Italian.

These traces demonstrate that a small model + KG agent outperforms frontier models (Gemini 2.5 Flash) on complex kinship reasoning — a key finding of the Hugging Face Build Small Hackathon 2026.

What's in a trace?

Each record contains:

Field	Description
scenario	The kinship graph structure (nodes & edges)
narrative	Italian-language description of the family situation
question	The inheritance-law question to answer
gold_answer	The correct answer
agent_trace	Full sequence of tool calls + LM reasoning steps
result	Final agent answer and correctness

Languages

• Italian 🇮🇹 — names, narratives, and legal scenarios
• English 🇬🇧 — ontology terms, tool-call schemas, and reasoning

Dataset Structure

Format

JSON Lines (.jsonl) — one trace per line.

Example

{
  "scenario": {
    "nodes": ["Marco", "Lucia", "Paolo"],
    "edges": [
      {"from": "Marco", "to": "Lucia", "relation": "spouse"},
      {"from": "Marco", "to": "Paolo", "relation": "parent"}
    ]
  },
  "narrative": "Marco e Lucia sono sposati. Paolo è il figlio di Marco...",
  "question": "Chi eredita i beni di Marco in assenza di testamento?",
  "gold_answer": "Lucia (coniuge) e Paolo (figlio) ereditano in parti uguali.",
  "agent_trace": [
    {"step": 1, "type": "tool_call", "tool": "lookup_relationship", "args": {"subject": "Marco", "predicate": "hasSpouse"}, "result": "Lucia"},
    {"step": 2, "type": "tool_call", "tool": "lookup_relationship", "args": {"subject": "Marco", "predicate": "hasChild"}, "result": "Paolo"},
    {"step": 3, "type": "reasoning", "content": "Marco has a surviving spouse (Lucia) and one child (Paolo). Under Italian succession law..."}
  ],
  "result": {"answer": "Lucia e Paolo ereditano in parti uguali", "correct": true}
}

Intended Uses

• Training data for SFT of small models on KG-grounded reasoning
• Evaluation of tool-use agents on multi-hop kinship queries
• Analysis of agent reasoning patterns and failure modes
• Benchmarking ontology-aware LM agents

Source & Generation

Traces were generated by the Sangue e Grafi agent pipeline:

1. Adversarial kinship scenarios generated with controlled graph complexity
2. Agent (Gemma 4B SFT+GRPO) solves each scenario using OWL ontology-grounded tool calls
3. Each trace is recorded with full tool-call history and reasoning chain

Associated Models

Model	Description
Gemma 4B SFT	SFT adapter trained on teacher traces
Gemma 4B GRPO	GRPO adapter (RL stage)
Nemotron 4B SFT	Cross-architecture SFT baseline
Nemotron 4B GRPO	Cross-architecture GRPO
GGUF (quantized)	Ready-to-run quantized model

Project Links

Resource	Link
🚀 Live Demo	HF Space
📦 GitHub	cyberandy/sangue-e-grafi
📄 Paper	RLM-on-KG (arXiv:2604.17056)

Citation

@misc{sangue-e-grafi-2026,
  title   = {Sangue e Grafi: Small Models Beat Frontier LLMs on Adversarial Kinship Reasoning with Knowledge Graph Agents},
  author  = {Andrea Volpini},
  year    = {2026},
  url     = {https://github.com/cyberandy/sangue-e-grafi},
  note    = {Hugging Face Build Small Hackathon 2026}
}