Training scripts grow into 500 lines of boilerplate that nobody can reproduce. Axolotl puts every training decision into one YAML: data, quantization, LoRA rank, packing, parallelism. Here is a minimal config that runs.

📜 Table of Contents

• Who this fits
• A QLoRA config that actually runs
• Install and run
• Data format: type decides parsing
• Does it work? Verify in four steps
• Axolotl or something else
• FAQ
• Related

Axolotl Fine-Tuning Tutorial 2026: Config-Driven LLM Training With Real Code

A training script grows into 500 lines of boilerplate: data loading, chat templates, quantization, gradient accumulation, checkpoints, logging — rewritten for every model, with no guarantee two runs match.

Axolotl puts every decision into one YAML. LoRA? A config line. 4-bit? Two lines. Sequence packing, gradient checkpointing, chat templates, DeepSpeed — all config. The benefit is not brevity: the whole recipe goes into Git, diffs, rolls back, and survives code review.

Who this fits

Teams that need reviewable, reproducible training; multi-GPU, FSDP and DeepSpeed setups; hyperparameter sweeps by editing YAML. For a single-GPU one-off LoRA, Unsloth is faster to start.

A QLoRA config that actually runs

base_model: Qwen/Qwen2.5-7B-Instruct
load_in_4bit: true

adapter: lora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true

datasets:
  - path: data/train.jsonl
    type: alpaca
val_set_size: 0.05

sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true

micro_batch_size: 2
gradient_accumulation_steps: 4
num_epochs: 1
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 2e-4

gradient_checkpointing: true
flash_attention: true
output_dir: ./out

sample_packing is the most overlooked performance switch. Instruction samples are often a few hundred tokens; without packing, most of a 2048-token sequence is padding and the GPU computes nothing useful. The side effect is a variable number of samples per step, which changes how you read the loss curve.

Install and run

git clone https://github.com/axolotl-ai-cloud/axolotl
cd axolotl && pip install -e .
pip install -U flash-attn --no-build-isolation

axolotl preprocess config.yml
accelerate launch -m axolotl.cli.train config.yml

Merge afterwards:

axolotl merge-lora config.yml
# older versions: python -m axolotl.cli.merge_lora config.yml

Data format: type decides parsing

For multi-turn, prefer chat_template with train_on_inputs: false so loss is computed on answers only.

Data quality beats format. 500-5000 self-consistent samples beat tens of thousands of inconsistent ones. Label 100, train one epoch, read the output, then label the rest.

Does it work? Verify in four steps

1. Hold out 50-100 samples and compare base vs tuned on the same prompts.
2. Check the loss trend; below 0.1 usually means memorization.
3. Rerun 20 in-dataset prompts; watch for verbatim parroting.
4. Ask a few unrelated questions to catch general-ability regression.

Axolotl or something else

FAQ

Unknown config field? The schema changed across minor versions. Compare against the matching YAML under examples/ in the repo rather than old blog posts.

OOM? Lower sequence_len, then micro_batch_size with higher accumulation, confirm load_in_4bit, and only then reduce lora_r.

Resume after a crash? --resume_from_checkpoint ./out/checkpoint-500. Keep the checkpoint directory intact.

How many tokens and steps? The startup log prints sample count, packed step count and effective batch — check those before reading loss.

Multimodal models? Some are supported with a different config shape; use the official multimodal example.

Note: config fields evolve by version. Examples reflect the common 2026 structure; verify against the repo's examples/ and official docs.

Related

• LLM微调工具包对比：Unsloth vs Axolotl vs TRL
• Unsloth QLoRA教程2026：8GB显存就能跑的4-bit微调
• Unsloth微调教程：用LoRA训练自己的AI模型