Fine-Tune Any LLM Without Writing Training Code: HuggingFace AutoTrain in 5 Minutes

HuggingFace AutoTrain lets you SFT, DPO, or ORPO any open model in a web UI or single autotrain command — no training code, no server, adapter pushed to Hub automatically.

What you'll be able to do after this:

• Launch a supervised fine-tuning (SFT), DPO, or ORPO run on any open LLM without writing training code
• Prepare your dataset as JSONL and configure LoRA hyperparameters through AutoTrain's UI or a single CLI command
• Have the fine-tuned LoRA adapter automatically pushed to your HuggingFace Hub repo when training finishes

AutoTrain Advanced is HuggingFace's hosted fine-tuning platform. Pick a model, upload data, configure a few sliders — and it writes and runs the training loop for you. No SFTTrainer boilerplate, no CUDA debugging, no manual Hub upload.

Step 1: Prepare your dataset

AutoTrain expects JSONL with one training example per line. For SFT:

{"prompt": "Classify the sentiment of this review:", "completion": "Positive"}
{"prompt": "Summarize in one sentence:", "completion": "A bear visited a bakery."}

For DPO, each row needs prompt, chosen, and rejected fields.

Step 2: Create a project

Go to huggingface.co/autotrain → Create new project → LLM → pick your base model (e.g. Qwen/Qwen2.5-7B-Instruct or meta-llama/Llama-3.2-3B-Instruct) → upload JSONL → pick hardware (ZeroGPU is free for small runs) → Start Training.

AutoTrain provisions compute, installs the right CUDA stack, and streams training logs in the Space UI.

Or use the CLI (works locally or in CI):

pip install autotrain-advanced

autotrain llm \
  --train \
  --model "Qwen/Qwen2.5-3B-Instruct" \
  --data-path ./data \
  --trainer sft \
  --lora-r 16 \
  --lora-alpha 32 \
  --lr 2e-4 \
  --batch-size 2 \
  --gradient-accumulation 4 \
  --num-train-epochs 3 \
  --push-to-hub \
  --username your-hf-username \
  --token $HF_TOKEN

When training ends the LoRA adapter lands in your Hub repo — ready to merge with peft.merge_and_unload() or serve directly via AutoTrain's inference endpoint.

Safe starting defaults:

• --lora-r 16, --lora-alpha 32: balanced expressivity vs. adapter size
• --lr 2e-4: standard for QLoRA SFT
• --batch-size 2 + --gradient-accumulation 4: effective batch size 8, fits on a 16GB GPU

Sources: AutoTrain LLM Fine-tuning — HuggingFace Docs · Low-Code LLM Alignment with AutoTrain — HuggingFace Blog