⚖️ Fine-Tuning Indonesian Legal QA with GRPO (In Progress)
This repository documents the real-world fine-tuning process of an Indonesian legal case QA dataset using Group Relative Policy Optimization (GRPO) — a reinforcement learning technique optimized for language models.
Rather than a "proof of concept," this is a proof of practice:
✅ That fine-tuning legal QA models is achievable
✅ That free tools like Kaggle and Hugging Face are enough
✅ That you don’t need a paid GPU or local compute to build powerful NLP systems
🔄 Fine-Tuning Workflow
✅ Model: Lightweight, efficient transformer-based model (ideal for legal QA)
✅ Pipeline: Cloud-to-cloud push/pull of checkpoints via Hugging Face Hub
✅ Epochs: Trained across 3 full epochs on the complete dataset
Checkpoints are updated incrementally and stored in Hugging Face for easy access and reproducibility.
🛠️ Current Status
🟡 Training in progress
📈 GRPO working efficiently within Kaggle limits
⏱️ Approx. 72 hours projected for full run (within Kaggle’s free GPU quota)
✅ All training, saving, and transfer operations are done via cloud tools — no local GPU needed
🎯 Why This Matters
This repository is a proof that:
🧑🎓 Students and researchers can fine-tune advanced models using only free cloud tools
🧪 GRPO can be applied in resource-constrained environments
📚 Legal NLP is now more accessible than ever
🌍 Anyone, from anywhere, can participate in open AI development — no expensive infrastructure required
📦 What’s Inside
🧠 Indonesian legal QA dataset fine-tuned with GRPO
📊 Checkpoints from each training epoch
🔁 Compatible with transformers, peft, and RL libraries
📝 Simple logs and setup for reproducibility
⚠️ Disclaimer
This repository is for learning, research, and development only.
The dataset is derived from public consultations and may not represent finalized legal interpretations.
