Description
I’m building an AI VTuber project and need help specifically with the LLM fine-tuning/data pipeline. The target model is Mistral NeMo 12B or a similar 12B–14B open-source model using LoRA/QLoRA. The project needs to run on a 24GB GPU alongside other services like vision and voice/TTS, so efficiency matters. This is not a generic chatbot project. The model needs to learn structured VTuber interaction behavior, including: * Twitch chat responses * streamer/voice responses * screen/game reactions * proactive riffs * silence/ignore behavior * callback/continuity behavior * viewer TTS/high-priority chat messages I already have an existing project, dataset attempts, and some cleanup work started, but I need someone experienced to review the current setup and help build a proper training pipeline. What I need help with: 1. Audit the current dataset and fine-tuning setup. 2. Clean, deduplicate, and structure training examples. 3. Build a high-quality Tier A dataset instead of using noisy random stream fragments. 4. Prepare or fix LoRA/QLoRA training scripts. 5. Run a small test fine-tune. 6. Build a simple A/B evaluation to compare the fine-tuned model against the current prompt/guard system. 7. Document the pipeline so I can rerun it later. Please include examples of previous LLM fine-tuning work, especially anything involving Mistral, LoRA/QLoRA, chatbots, AI agents, Twitch, Discord, VTubers, or real-time conversational systems.