Slotherynn/legal-chatbot-qwen-sft
Slotherynn/legal-chatbot-qwen-sft is a supervised fine-tuning experiment based on Qwen2.5-3B-Instruct, developed by Slotherynn. This model was trained using 4-bit QLoRA on a general Indonesian instruction dataset, not a specialized legal QA dataset. Despite its name, it is not intended as an Indonesian legal expert and failed to produce valid grounded outputs in a legal RAG benchmark. It serves primarily as a learning artifact for fine-tuning methods.
Loading preview...
What is this model about?
This model, Slotherynn/legal-chatbot-qwen-sft, is a supervised fine-tuning (SFT) experiment built upon the Qwen2.5-3B-Instruct base model. It was developed by Slotherynn as a learning artifact to explore fine-tuning techniques, specifically using 4-bit QLoRA with Unsloth and TRL.
What makes THIS different from all the other models?
Unlike many models, this one is explicitly documented as a failed experiment for its intended production use case. While its name suggests legal expertise, it was trained on a general Indonesian instruction dataset (Ichsan2895/alpaca-gpt4-indonesian) rather than curated legal data. A legal RAG benchmark showed it produced zero valid grounded outputs and could not reliably follow citation or abstention contracts, indicating it is not suitable for legal applications. This transparency about its limitations and experimental nature sets it apart.
Should I use this for my use case?
No, you should NOT use this model for legal advice, production legal answers, or any application requiring reliable legal expertise.
Key Limitations:
- Not a legal expert: It was not trained on legal-specific datasets and does not understand legal nuances.
- Failed benchmark: It performed poorly in a legal RAG benchmark, failing to generate grounded outputs or adhere to citation requirements.
- Historical data: The benchmark corpus is historical and does not reflect current law.
- No reliable citation/abstention: It cannot reliably follow required output contracts.
This model is best suited as a learning artifact for developers interested in observing the outcomes of fine-tuning Qwen2.5-3B-Instruct on general instruction data, particularly regarding its limitations when applied to specialized domains without appropriate training data.