legal-chatbot-qwen-grpo API & Inference Endpoint

Training

Method: GRPO with Unsloth and TRL after the historical SFT run.
Dataset: first 2,000 examples from Ichsan2895/alpaca-gpt4-indonesian, licensed CC BY-SA 4.0 and derived from FreedomIntelligence/alpaca-gpt4-indonesian.
LoRA: rank 16, alpha 16, applied to attention and MLP projection modules.
Maximum sequence length: 1024.
Training: 200 steps, learning rate 5e-6.
Historical rewards: format, reasoning length, ROUGE-L correctness, and language. Training logs show the format and reasoning rewards remained zero while the language reward remained constant, so these rewards did not provide the intended learning signal.

Legal RAG Benchmark

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.

Table with columns: Metric, Result
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.

Limitations and Intended Use

Not for production legal answers or legal advice.
Not trained on the four regulatory documents used by the benchmark.
Does not reliably follow the required citation and abstention contract.
The reward design is unsuitable evidence of grounded legal reasoning.
The benchmark corpus is historical and does not represent current law.

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.

Training

Method: GRPO with Unsloth and TRL after the historical SFT run.
Dataset: first 2,000 examples from Ichsan2895/alpaca-gpt4-indonesian, licensed CC BY-SA 4.0 and derived from FreedomIntelligence/alpaca-gpt4-indonesian.
LoRA: rank 16, alpha 16, applied to attention and MLP projection modules.
Maximum sequence length: 1024.
Training: 200 steps, learning rate 5e-6.
Historical rewards: format, reasoning length, ROUGE-L correctness, and language. Training logs show the format and reasoning rewards remained zero while the language reward remained constant, so these rewards did not provide the intended learning signal.

Legal RAG Benchmark

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.

Table with columns: Metric, Result
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

Limitations and Intended Use

Not for production legal answers or legal advice.
Not trained on the four regulatory documents used by the benchmark.
Does not reliably follow the required citation and abstention contract.
The reward design is unsuitable evidence of grounded legal reasoning.
The benchmark corpus is historical and does not represent current law.

legal-chatbot-qwen-grpo

README

Training

Legal RAG Benchmark

Limitations and Intended Use

Explore FriendliAI today

README

Training

Legal RAG Benchmark

Limitations and Intended Use