Fine-Tuning.

Fine-tune models in minutes

Generate training data, dispatch jobs to four providers, and deploy serverless models — all in under 18 minutes of hands-on work.

macOS · Windows · Linux

Fine-Tuning in Kiln

Training data without a data team

Kiln synthetic data generation can build a fine-tuning dataset in minutes. Topic trees generate diverse samples and prevent near-duplicates. Start from a few examples, scale to 1,000+ with our ladder strategy, and curate interactively before training.

Train anywhere, in a few clicks

Train on Fireworks, Together, OpenAI or Google Vertex in a few clicks — without managing GPUs. Or download datasets for training on any infrastructure.

Over 60 fine-tuneable models

Choose from over 60 models including Qwen, DeepSeek, GLM, Llama, GPT, Gemini and more. Don't worry about training dataset formats or tokenizers — it's handled by Kiln. Tune all weights, or LoRAs.

Deploy serverless or export weights

Models are deployed automatically, and scale to zero when not in use. Fine-tuned models, with the convenience of pay-per-token. Or export weights and deploy anywhere.

Distillation made easy

Kiln can distill large state of the art models into smaller, faster and cheaper models. Reduce costs, increase speed, and maintain intelligence.

Output, tools & reasoning

Train for everything that drives quality: output, reasoning structure and tool use.

Fine-tuning before and after Kiln

Without Kiln
  • Write data formatting scripts for each provider, debug JSONL schemas, and manage training configs in notebooks.
  • Training data takes weeks to generate and curate.
  • Custom code, data formats, and tokenizer work for every model you want to tune.
With Kiln
  • Pick a model, select a dataset, and click Train. Kiln handles provider formats, deployment, and configs.
  • Training datasets built in minutes with synthetic data generation, and topic trees to ensure diversity and coverage.
  • Dispatch dozens of training jobs across models and providers without coding or dataset formatting. Find the best with structured evals.

Frequently asked

Where does fine-tuning run?

Choose your provider: OpenAI, Google Gemini, Fireworks, or Together AI. Or export datasets to train any model on your own infra with Unsloth, Axolotl, or Colab.

How much does fine-tuning cost?

It varies, but in our demo we train 5 Llama models + Mixtral on Fireworks for $1.47 and GPT-Mini for $2.03. Fine-tunes deploy serverless with no recurring costs.

Does my training data leave my machine?

Kiln is local-first. Data is sent to provider APIs only when you dispatch a job. Export and train entirely on your own infra if data residency requires it.

Can I fine-tune for tool calling or reasoning?

Yes to both. Reasoning distillation trains smaller models on thinking traces from large models like DeepSeek or Claude. Tool-use training teaches a model your tool set, with a built-in eval to measure improvement.

Ship a fine-tuned model before lunch.

Generate training data, train across four providers, deploy serverless, verify with evals.