Downloads · 30 days
9
2% of all-time downloads
Slotherynn/legal-chatbot-qwen-grpo
legal-chatbot-qwen-grpo is a text generation model from Slotherynn. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
This is a historical GRPO experiment derived from Slotherynn/legal-chatbot-qwen-sft. Despite its repository name, it was trained on general Indonesian instructions rather than provenance-backed legal QA data and must…
Downloads · 30 days
9
2% of all-time downloads
All-time downloads
391
Public
Parameters
3.1B
6.2 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors6.2 GB · 100%
From the Hugging Face model README
This is a historical GRPO experiment derived from
Slotherynn/legal-chatbot-qwen-sft. Despite its repository name, it was
trained on general Indonesian instructions rather than provenance-backed legal
QA data and must not be treated as an Indonesian legal expert.
Ichsan2895/alpaca-gpt4-indonesian, licensed CC BY-SA 4.0 and derived from
FreedomIntelligence/alpaca-gpt4-indonesian.5e-6.The model was evaluated on 60 reviewed Indonesian legal cases using the same frozen retrieval results, prompt, deterministic decoding, and validator as the base and SFT models.
| Metric | Result |
|---|---|
| Attempted generations | 43 |
| Valid grounded outputs | 0 |
| Citation precision | Not evaluable |
| Faithfulness | Not evaluable |
| Answer relevance | Not evaluable |
| Abstention accuracy | 0.2167 |
| Mean generation latency | 12.15 s |
| Peak VRAM (4-bit load) | 2686.36 MB |
| Loaded footprint | 1916.96 MB |
Every attempted generation failed the required grounded-output contract. GRPO did not improve any primary quality metric over either the base or SFT model, so this model is rejected for the production RAG application.
The project retains this model and its training notebook only as a learning artifact. Any future GRPO run must use legal QA data with clear provenance and rewards for groundedness, citation correctness, abstention, and answer quality.