kimeunsur/llama3-8b-trex-triples

This repository provides a knowledge-graph–oriented variant of Meta-Llama-3-8B-Instruct, fine-tuned with LoRA on the T-REx factual triple dataset as a triple extraction model.

The model takes natural language passages as input and outputs factual (head, relation, tail) triples in a structured JSON format, designed to be used as a knowledge graph backend for RAG systems (e.g., HippoRAG-style KG retrieval).

---

Training Details

Data

• Source: T-REx factual triple dataset
• Preprocessing (conceptual):
  • T-REx triples (subject, relation, object) were converted into simple natural language sentences.
  • The model was supervised to recover the triples in a structured JSON format:

    {
      "triples": [
        ["New Bomb Turks", "P136", "garage punk"]
      ]
    }
    

  • Relations are expressed either as short phrases (e.g. "is from") or as Wikidata-style relation IDs (e.g. "P136").