Training and Fine-Tuning Recipes

19 resources

• Gemma 3 1B instruction tuning with QLoRA

  A compact Axolotl configuration for instruction tuning a 4-bit Gemma 3 checkpoint with trainable adapters.

  Axolotl · Public publisher configuration; model and dataset terms apply.

  Dataset

  cgato/SlimOrcaDedupCleaned

  Framework

  Axolotl

  Hardware

  Quantized training configuration. Minimum VRAM unspecified.

• Llama 3 8B preference tuning with LoRA and DPO

  An Axolotl recipe that trains Llama 3 adapters from preferred and rejected chat responses.

  Axolotl · Public publisher configuration; model and dataset terms apply.

  Dataset

  fozziethebeat/alpaca_messages_2k_dpo_test

  Framework

  Axolotl

  Hardware

  Minimum VRAM unspecified.

• Train a Qwen3 reward model on preference pairs

  A runnable example that trains a scalar reward scorer using the Skywork Qwen3 checkpoint.

  Axolotl · Public publisher configuration; model and dataset terms apply.

  Dataset

  argilla/distilabel-intel-orca-dpo-pairs

  Framework

  Axolotl

  Hardware

  Minimum VRAM unspecified.

• Fine-tune Orpheus 3B for speech generation

  A documented training recipe for adapting an Orpheus text-to-speech checkpoint to a transcribed audio dataset.

  Axolotl · Public publisher guide and configuration; bring your own dataset.

  Dataset

  User-supplied transcribed audio, prepared as documented.

  Framework

  Axolotl

  Hardware

  Publisher documents a NVIDIA GPU with at least 45 GB VRAM for the supplied settings.

• Qwen2.5-VL 7B image-and-text instruction tuning

  A LoRA training configuration for a Qwen vision-language model using image and conversation pairs.

  Axolotl · Public publisher configuration; model and dataset terms apply.

  Dataset

  HuggingFaceH4/llava-instruct-mix-vsft

  Framework

  Axolotl

  Hardware

  Minimum VRAM unspecified.

• GPT-OSS 20B full fine-tuning across GPUs with FSDP2

  A distributed supervised fine-tuning configuration for updating GPT-OSS 20B model weights.

  Axolotl · Public publisher configuration and documented launch command.

  Dataset

  HuggingFaceH4/Multilingual-Thinking

  Framework

  Axolotl

  Hardware

  Publisher examples use 8×48 GB GPUs or 4×80 GB GPUs for this configuration.

• Gemma 3 12B quantization-aware training for NVFP4

  A Gemma 3 configuration that includes NVFP4 weight and activation quantization in the training setup.

  Axolotl · Public publisher configuration; model and dataset terms apply.

  Dataset

  tatsu-lab/alpaca

  Framework

  Axolotl

  Hardware

  Minimum VRAM unspecified. Verify NVFP4 support in your stack.

• Llama 3.1 long-context tuning with ring context parallelism

  A multi-GPU recipe for distributing long training sequences across GPUs with ring context parallelism.

  Axolotl · Public publisher configuration with launch instructions.

  Dataset

  tatsu-lab/alpaca

  Framework

  Axolotl

  Hardware

  Eight GPUs and PyTorch 2.13 or newer per the example. Minimum VRAM unspecified.

• Distill Qwen2.5 14B into a Qwen2.5 7B adapter

  A single-device teacher–student recipe that uses Qwen2.5 14B guidance to train a 7B student adapter.

  PyTorch torchtune · Public publisher configuration; local checkpoints required.

  Dataset

  Alpaca Cleaned via torchtune.datasets.alpaca_cleaned_dataset.

  Framework

  PyTorch torchtune

  Hardware

  Single CUDA device in the configuration. Minimum VRAM unspecified.

• Gemma 2 2B distributed full fine-tuning

  A torchtune example for updating the full Gemma 2 2B checkpoint on multiple CUDA GPUs.

  PyTorch torchtune · Public publisher configuration; local checkpoint required.

  Dataset

  Alpaca Cleaned via torchtune.datasets.alpaca_cleaned_dataset.

  Framework

  PyTorch torchtune

  Hardware

  The publisher requires at least two GPUs for this configuration and shows a four-device CUDA launch.

• Train a WikiSQL LoRA adapter on Apple silicon

  An MLX LM configuration for fine-tuning a small Llama model on structured query examples on a Mac.

  MLX LM · Public publisher configuration; model and dataset terms apply.

  Dataset

  mlx-community/WikiSQL

  Framework

  Apple MLX LM

  Hardware

  Apple silicon. Minimum unified memory unspecified.

• Train visual math reasoning with GRPO

  A TRL example for improving image-based math answers with automatically checked rewards.

  Hugging Face TRL · Public publisher script and launch examples.

  Dataset

  lmms-lab/multimodal-open-r1-8k-verified

  Framework

  Hugging Face TRL

  Hardware

  Minimum VRAM unspecified.

• Teach Qwen3 to play Sudoku with GRPO and OpenEnv

  A reinforcement-learning recipe for training a model to make valid moves over multiple interactions with Sudoku.

  Hugging Face TRL · Public publisher script and notebook; a configured OpenEnv endpoint is required.

  Dataset

  Interactive Sudoku games from the OpenEnv TextArena environment.

  Framework

  Hugging Face TRL

  Hardware

  One GPU with colocated vLLM or two GPUs for the separate-server example. Publisher tested on one A100 80 GB.

• Distill a chat model with TRL's GOLD trainer

  An on-policy distillation example that trains a smaller chat model using a separate teacher.

  Hugging Face TRL · Public publisher script and launch examples.

  Dataset

  nvidia/Nemotron-Pretraining-Specialized-v1.1

  Framework

  Hugging Face TRL

  Hardware

  Minimum VRAM unspecified.

• Tune vision-language responses with grounded preference pairs

  A multimodal DPO training script for comparing grounded and ungrounded answers about images.

  Hugging Face TRL · Public publisher script and launch examples.

  Dataset

  HuggingFaceH4/rlaif-v_formatted

  Framework

  Hugging Face TRL

  Hardware

  Minimum VRAM unspecified.

• Fine-tune EmbeddingGemma for dense retrieval

  An Unsloth example for adapting a compact embedding model to retrieve relevant passages.

  Unsloth · Public publisher notebook export; model and dataset terms apply.

  Dataset

  tomaarsen/miriad-4.4M-split

  Framework

  Unsloth

  Hardware

  The exported example uses CUDA. Minimum VRAM unspecified.

• Adapt Gemma 3n E4B for German audio transcription

  An audio-input training example using Gemma 3n and a German speech dataset.

  Unsloth · Public publisher notebook export; model and dataset terms apply.

  Dataset

  kadirnar/Emilia-DE-B000000

  Framework

  Unsloth

  Hardware

  The exported example uses CUDA. Minimum VRAM unspecified.

• Personalize a diffusion model with DreamBooth LoRA

  An official Diffusers workflow for training a small image-generation adapter from photos of a specific subject.

  Hugging Face Diffusers · Public publisher script and examples; checkpoint and image rights apply.

  Dataset

  diffusers/dog-example

  Framework

  Hugging Face Diffusers

  Hardware

  Publisher documents memory-saving settings. Minimum VRAM for the LoRA example is unspecified.

• Train a ControlNet on paired conditioning images

  A Diffusers implementation of ControlNet training for teaching an image model to follow an auxiliary image condition.

  Hugging Face Diffusers · Public publisher script and examples; checkpoint and dataset terms apply.

  Dataset

  fusing/fill50k

  Framework

  Hugging Face Diffusers

  Hardware

  Publisher reports about 38 GB VRAM for the default setup and documents lower-memory alternatives.