Axolotl Fine-Tuning

• You are writing or debugging Axolotl YAML configs and want configuration-aware guidance
• You are setting up LoRA, QLoRA, or full fine-tuning runs on Llama, Mistral, or other supported architectures
• You are implementing preference alignment (DPO, KTO, ORPO, GRPO) and need help with dataset format and training setup
• You are scaling to multi-GPU training with FSDP v2 or DeepSpeed and need help with parallelism settings

• You need a hosted fine-tuning API — Axolotl is a self-hosted, local training framework requiring your own GPU hardware
• You are working with frameworks other than Axolotl (e.g. LLaMA-Factory, Unsloth) — skill content is Axolotl-specific
• You want inference or deployment help without any training involved

The skill covers Axolotl's full YAML config surface — model selection, dataset paths, training hyperparameters, and advanced options like save_compressed and context_parallel_size. Get accurate, config-aware answers without digging through docs.

Get detailed help configuring LoRA adapters and quantized QLoRA setups via PEFT. The skill includes guidance on rank, alpha, target modules, and common pitfalls when combining quantization with adapter training.

Covers DPO, KTO, ORPO, and GRPO — including dataset format requirements, trainer configuration, and reward modeling considerations for each method.

Includes FSDP v2 and DeepSpeed configuration patterns, context parallelism setup, and guidance on running NCCL tests to validate multi-GPU communication before launching large jobs.

Comprehensive coverage of Axolotl's supported dataset formats and prompt strategies, including RawInputOutputPrompter and batched vs. single-example data handling.

Guidance extends to multimodal (vision-language) model architectures supported by Axolotl, helping you configure training for models that handle both text and image inputs.

Configure a LoRA adapter run on a Llama-family model — select target modules, set rank and alpha, and wire up your dataset in the YAML config with assistant guidance throughout.

Set up a DPO training run from preference-labeled data: format the dataset correctly, configure the trainer, and tune hyperparameters with accurate, doc-grounded guidance.

Configure FSDP v2 with TRANSFORMER_BASED_WRAP and context parallelism, or set up DeepSpeed offloading — the skill covers the YAML fields and trade-offs for each approach.

Diagnose issues like sequence length overflow, data collation mismatches, or NCCL bottlenecks using patterns and utilities documented in the skill's reference files.