Skip to main content

​
HTTP API: Start a Fine-Tune

You can start a fine-tune directly against a repository via the HTTP API. Endpoint
  • POST /api/repos/{owner}/{repo}/fine_tunes/
Path Parameters
  • owner – Namespace or username that owns the repo.
  • repo – Repository name.
Request Body
  • resource – Path to the dataset within the repository (for example main/suriya-images/images.parquet).
  • base_model – Base model identifier to fine-tune (for example meta-llama/Llama-3.2-1B-Instruct).
  • script_type – Type of training script to run (for example text_generation).
  • training_params – Object containing training hyperparameters such as:
    • question_column – Column in the dataset to use as the input/question.
    • answer_column – Column in the dataset to use as the label/answer.
    • batch_size, epochs, learning_rate, logging_steps, seq_length, etc.
    • LoRA-related options like use_lora, lora_alpha, lora_rank.
Below is an example of starting a fine-tune for a text generation model: 
curl --location 'http://localhost:3001/api/repos/surya/random/fine_tunes/' \
--data '{
    "resource": "main/suriya-images/images.parquet",
    "training_params": {
      "answer_column": "file_path",
      "batch_size": 1,
      "enable_thinking": false,
      "epochs": 1,
      "grad_accum": 1,
      "learning_rate": 0.0001,
      "logging_steps": 10,
      "lora_alpha": 16,
      "lora_rank": 16,
      "neftune_noise_alpha": 0,
      "question_column": "file_path",
      "save_steps_ratio": 0.25,
      "save_strategy": "epoch",
      "seq_length": 1024,
      "use_lora": true
    },
    "base_model": "meta-llama/Llama-3.2-1B-Instruct",
    "script_type": "text_generation"
  }'