🚧 Gemma 2 Facts (Work in Progress)
Status: Active Training.
Current State: Synthetic data generation pipeline is running.
Target: Fine-tuning Gemma 2 (9B/2B) for strict factual grounding using Direct Preference Optimization (DPO).
Project Goal
This model is being trained to solve the "hallucination" problem in RAG systems. It is aligned to refuse questions that cannot be answered strictly from the provided context document.
- Dataset: Custom synthetic expansion of the DeepMind FACTS Grounding benchmark.
- Method: DPO + QLoRA on consumer hardware (T4 GPU).
