本项目利用医学领域的 CoT 数据对 Deepseek-R1-Distill-Qwen-7B 进行微调，通过 QLoRA 量化和 Unsloth 加速训练，显著提升模型在复杂医学推理任务中的慢思考能力。知识蒸馏技术使轻量级模型获得大模型的推理优势，实现高效、准确且具有解释性的医学问答系统。

⚡ Enterprise-ready federated learning for LLMs. On-device personalization, privacy-preserving aggregation, LoRA adapters, real-time dashboard. Perfect for privacy-sensitive AI applications. 🌟🏆

A 4-bit NormalFloat (NF4) quantized version of InstaDeep's Nucleotide Transformer 2.5B Multi-Species model, optimized for significantly lower GPU memory consumption while preserving embedding quality and inference performance.

A collection of Jupyter notebooks for fine-tuning small language models using LoRA and 4-bit quantization across multiple architectures (Phi, Gemma, DialoGPT, DeepSeek, SmolLM).

Complete setup guide for Ollama + Gemma 4 E4B + Hermes Agent on a 16 GB Apple Silicon iMac — automated install script, fine-tuned config, and troubleshooting

Fine-Tuning Mistral-7B with Unsloth is a streamlined implementation for efficiently adapting the powerful Mistral-7B language model using the Unsloth framework. This project showcases low-rank adaptation (LoRA), 4-bit quantization, and structured conversational datasets to fine-tune large models with minimal memory overhead and maximum performance.

QLoRA-enhanced Qwen2.5-3B model with 4-bit quantization for AI research Q&A.

End-to-end healthcare LLM system using Phi-2 + QLoRA (4-bit quantization) with FastAPI inference API.

Fine-tune transformer models for IMDb sentiment classification using QLoRA (4-bit quantization) and PEFT. Built with Hugging Face Transformers, BitsAndBytes, and FastAPI for efficient training, inference, and deployment of large language models.

이미지·상황·질문에서 정답을 고르되 근거가 부족하면 '모름'을 택하도록 학습시킨 멀티모달 편향 QA — 성균관대 챌린지