Instruction dataset
        ↓
Fine-tuned GPT-2 Medium
        ↓
Generated responses
        ↓
Judge LLM
        ↓
Quality scores

        __
        __