siwo/govon-legal-adapter-train

GovOn Legal Adapter Training

EXAONE 4.0-32B + Unsloth QLoRA fine-tuning for Korean legal domain.

Requirements

• —Hardware: GPU Space required (A100 80GB recommended)
• —Secret: HF_TOKEN with write access

What it does

1. 1.Loads EXAONE 4.0-32B in 4-bit quantization
2. 2.Applies LoRA (r=16, alpha=32) to attention + MLP layers
3. 3.Trains on 50K samples from neuralfoundry-coder/korean-legal-instruction-sample
4. 4.Pushes the adapter to siwo/govon-legal-adapter

Health endpoint

GET / returns JSON with training status (stage, step, loss). If no GPU is detected, status shows {"stage": "error", "error": "..."} for 10 minutes before the process exits.