---
license: apache-2.0
base_model: unsloth/Qwen3.5-0.8B
library_name: peft
pipeline_tag: text-generation
tags:
  - knowledge-graph
  - relation-extraction
  - entity-extraction
  - qlora
  - lora
  - sft
  - qwen
datasets:
  - Babelscape/rebel-dataset
language:
  - en
---

# tiny-graph-extractor-qwen3.5-0.8b-lora

A **LoRA adapter** for `unsloth/Qwen3.5-0.8B`, fine-tuned to extract entities and (subject, relation, object) triplets from sentences for knowledge graph ingestion.

This repository contains **adapter weights only**.

The goal of this project is to replace expensive frontier-LLM calls in the knowledge graph extraction pipeline of [rst0070/knowledge-base](https://github.com/rst0070/knowledge-base) with a tiny model that runs locally.

- **Base model**: `unsloth/Qwen3.5-0.8B` (Instruct)
- **Method**: QLoRA SFT (4-bit frozen base + LoRA adapters)
- **Training data**: [REBEL](https://huggingface.co/datasets/Babelscape/rebel-dataset) — Wikipedia sentences paired with Wikidata triplets
- **Hardware**: single 16GB GPU (RTX 4060 Ti)
- **Source**: [github.com/rst0070/tiny-entity-extractor](https://github.com/rst0070/tiny-entity-extractor)

## Known Contract: Wikipedia Style

The adapter was trained on REBEL, so its output follows a **Wikipedia-style contract**:

- Entity surface forms follow Wikipedia conventions
- Relations follow Wikidata definitions and granularity

This is the contract the model is built against. Inputs and evaluation should stay inside that contract — domain-specific terminology, informal text, or alternate relation ontologies are out of scope and will degrade quality.

## Output Format

The model emits structured JSON:

```json
{
  "entities": ["Entity A", "Entity B"],
  "relations": [
    {"head": "Entity A", "relation": "relation_type", "tail": "Entity B"}
  ]
}
```

## Usage

```python
from transformers import pipeline

pipe = pipeline(
    "text-generation", model="rst0070/tiny-graph-extractor-qwen3.5-0.8b-lora",
    max_new_tokens=1024,
)
messages = [
    {
        "role": "system",
        "content": (
            "You are a knowledge graph extraction assistant. "
            "Given a text, extract all entities and their relations as JSON. "
            "Output only valid JSON with no additional text."
        )
    },
    {
        "role": "user",
        "content": "Apple was founded by Steve Jobs, Steve Wozniak, and Ronald Wayne in 1976. Steve Jobs served as the CEO of Apple.",
    },
]

result = pipe(messages)
response = result[0]['generated_text'][-1]
print(response)

import re
import json

fence_re = re.compile(r"```(?:json)?\s*\n(.*?)\n```", re.DOTALL)
match = fence_re.search(response["content"])
if match:
    print(json.loads(match.group(1)))

```

## Training

The training target was constructed to deviate as little as possible from the base model's natural output, so SFT only has to close the smallest possible gap:

1. **Observe** how the base model formats answers on REBEL inputs with no fine-tuning.
2. **Align** the SFT target to match that observed format/phrasing where possible, while staying factually correct.
3. **Train** with QLoRA SFT against that aligned target.

## Limitations

- **English / Wikipedia distribution only.** Performance on other languages, domains (medical, legal, financial), or informal text is unknown and likely poor.
- **Wikidata relation ontology.** Relations outside the Wikidata vocabulary will not be produced reliably.
- **Sentence-level inputs.** Trained on REBEL sentence-level examples; not validated on long documents.
- **No factual grounding.** The model extracts what it reads; it does not verify claims.

## License

Apache 2.0, inheriting from the base model and dataset license terms.