Fine-tuning
Curated Fine-tuning tutorials.
AI Model Compression: Pruning, Quantization, and Knowledge Distillation
Three pillars of model compression: quantization (GPTQ/AWQ/GGUF, 8/4-bit), pruning, and knowledge distillation — principles, real tools, and accuracy/speed/memory trade-offs.
AdvancedAI Model Merging: SLERP, TIES, DARE, and Model Soup Techniques
Model merging fuses multiple fine-tuned models into one without retraining. Based on mergekit: linear/SLERP/TIES/DARE/task-arithmetic methods, real YAML configs, and when to use each.
AdvancedAI Model Quantization (GPTQ, AWQ): Complete Developer Guide 2026
A Complete Guide to AI Model Quantization (GPTQ/AWQ) (2026): Store weights with fewer bits to save VRAM and boost speed. Comparison of GPTQ vs AWQ, bitsandbytes/GGUF, selecting the 4-bit sweet spot, and a practical path of "directly downloading pre-quantized weights + deploying with vLLM/Ollama."
AdvancedFine-Tuning GPT-4o Mini: OpenAI Fine-Tuning API Complete Guide
GPT-4o Mini Fine-Tuning Complete Guide (2026): Use OpenAI's fine-tuning API to obtain a hosted model with stable format/style and reduce costs for massive calls. Includes real code for JSONL data preparation → upload → training → inference, when to fine-tune vs. prompt/RAG, and data quality > quantity.
AdvancedHugging Face SFT Trainer: Hands-On Tutorial
Hugging Face SFT Trainer Overview Supervised fine-tuning with Hugging Face TRL SFTTrainer. This tutorial provides a complete, runnable implementation. Prerequisites ```bash Install required packages pip install transformers datasets peft trl acce
IntermediateLarge Model Post-Training in Practice: From SFT to RL — The Complete Tech Stack
This article systematically explains the key techniques of large model post-training, including supervised fine-tuning (SFT), parameter-efficient fine-tuning (PEFT), reinforcement learning from human feedback (RLHF), and on-policy distillation (OPD). It focuses on the principles, pros and cons, and applicable scenarios of each method, and introduces a stability-plasticity trade-off framework to quantify the general capability loss caused by fine-tuning. By comparing the forgetting characteristics of full fine-tuning, LoRA, OFT, and other PEFT methods, it reveals that the destruction of activation space geometric structure is the key mechanism of forgetting. Finally, it summarizes the advantages of OPD as a new paradigm and provides practical guidelines and FAQs.
AdvancedLLM Fine-Tuning for Production: LoRA, QLoRA & RLHF in 2025
Fine-tuning LLMs allows adapting powerful foundation models to specific domains without training from scratch. This guide covers LoRA and QLoRA for parameter-efficient fine-tuning, dataset preparation and quality filtering, instruction tuning format, RLHF and DPO for alignment, fine-tuning on consumer GPUs with quantization, evaluation with domain benchmarks, and deploying fine-tuned models with vLLM or TGI for production serving.
AdvancedLLM Fine-tuning with LoRA: Complete Developer Guide 2026
A complete guide to fine-tuning large models with LoRA (2026): freeze the base model, train only low-rank adapters, finish in hours on a single GPU; QLoRA trains adapters on a 4-bit base. Includes real PEFT code, when to fine-tune (vs. prompting/RAG), and practical tips that data quality > quantity.
AdvancedFine-Tuning LLMs with LoRA and QLoRA: Complete Guide 2026
Complete guide to fine-tuning large language models using LoRA and QLoRA techniques in 2026. Covers dataset preparation, training configuration, hardware requirements, evaluation metrics, and deploying fine-tuned models to production.
AdvancedRLHF vs DPO: Training LLMs from Human Feedback - Technical Guide 2025
A guide to preference learning for alignment (2026): turning a base model into a helpful, harmless, honest assistant. RLHF (SFT + reward model + PPO) is complex but powerful; DPO uses a single preference loss, skipping the reward model and RL, making it simpler and more stable. Includes a comparison table and variants like IPO/KTO.