63 lines
1.8 KiB
Markdown
63 lines
1.8 KiB
Markdown
---
|
|
license: other
|
|
base_model: Qwen/Qwen2.5-0.5B-Instruct
|
|
language:
|
|
- ko
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- korean
|
|
- knowledge-graph
|
|
- qlora
|
|
- sft
|
|
- local-factual-data
|
|
---
|
|
|
|
# AIWORKX-KR-KG-Reasoner-0.5B-v0.10
|
|
|
|
## Overview
|
|
|
|
AIWORKX-KR-KG-Reasoner-0.5B-v0.10 is a small Korean-specialized language model fine-tuned from `Qwen/Qwen2.5-0.5B-Instruct`.
|
|
|
|
This model was created as a local MVP for K-AI leaderboard submission experiments. The training pipeline uses copyright-safe structured factual triples, converts them into a local knowledge graph, and generates Korean supervised fine-tuning data through deterministic rule-based templates.
|
|
|
|
## Base Model
|
|
|
|
- Base model: `Qwen/Qwen2.5-0.5B-Instruct`
|
|
- Fine-tuning method: QLoRA supervised fine-tuning
|
|
- Adapter merge: LoRA adapter merged into the base model
|
|
|
|
## Training Data
|
|
|
|
The training data was generated from local structured factual triples.
|
|
|
|
Pipeline:
|
|
|
|
1. Local factual triples
|
|
2. Knowledge graph nodes, edges, and evidence units
|
|
3. Rule-based Korean QA generation
|
|
4. SFT JSONL conversion
|
|
5. QLoRA fine-tuning
|
|
|
|
No OpenAI API, paid API, live external API, or web crawling was used in this MVP.
|
|
|
|
## Data Policy
|
|
|
|
The dataset uses structured factual triples rather than copied prose. The project is designed around copyright-safe factual data and knowledge-graph-based training.
|
|
|
|
## Intended Use
|
|
|
|
- Korean factual QA
|
|
- Knowledge-graph-grounded QA
|
|
- Triple-to-text generation
|
|
- Evidence-based relation extraction
|
|
- Small-scale K-AI leaderboard submission experiment
|
|
|
|
## Limitations
|
|
|
|
This is a small experimental model trained on a very small dataset. It may overfit to the rule-based format and may not perform well on broad open-domain tasks. It should not be used for high-stakes decisions.
|
|
|
|
## Version
|
|
|
|
- Version: 0.10
|
|
- Project motto: Copyright-safe factual data + Knowledge Graph + Local QLoRA
|