Fine Tuning • Synthetic Data Generation • Dataset Collaboration • Docs
| CI |        |
| Package |   |
| Meta |      |
| Apps |    |
- 🚀 Intuitive Desktop Apps: One-click apps for Windows, MacOS, and Linux. Truly intuitive design.
- 🎛️ Fine Tuning: Zero-code fine-tuning for Llama, GPT4o, and Mixtral. Automatic serverless deployment of models.
- 🤖 Synthetic Data Generation: Generate training data with our interactive visual tooling.
- 🤝 Team Collaboration: Git-based version control for your AI datasets. Intuitive UI makes it easy to collaborate with QA, PM, and subject matter experts on structured data (examples, prompts, ratings, feedback, issues, etc.).
- 📝 Prompt Generation: Automatically generate prompts from your data, including chain-of-thought, few-shot, and multi-shot, and more.
- 🌐 Wide Model and Provider Support: Use any model via Ollama, OpenAI, OpenRouter, Fireworks, Groq, AWS, or any OpenAI compatible API.
- 🧑💻 Open-Source Library and API: Our Python library and OpenAPI REST API are MIT open source.
- 🔒 Privacy-First: We can't see your data. Bring your own API keys or run locally with Ollama.
- 🗃️ Structured Data: Build AI tasks that speak JSON.
- 💰 Free: Our apps are free, and our library is open-source.
The Kiln desktop app is completely free. Available on MacOS, Windows and Linux.
In this demo, I create 9 fine-tuned models (including Llama 3.x, Mixtral, and GPT-4o-mini) in just 18 minutes, achieving great results for less than $6 total cost. See details.
Kiln is quite intuitive, so we suggest launching the desktop app and diving in. However if you have any questions or want to learn more, our docs are here to help.
- Quick Start
- How to use any AI model or provider in Kiln
- Fine Tuning LLM Models
- Synthetic Data Generation
- Collaborating with Kiln
- Rating and Labeling Data
- Prompt Styles
- Structure Data / JSON
- Organizing Kiln Datasets (Tags and Filters)
- Our Data Model
- Repairing Responses
- Keyboard Shortcuts
- Privacy Overview: Private by Design
For developers, see our Kiln Python Library Docs. These include how to load datasets into Kiln, or using Kiln datasets in your own code-base/notebooks.
Our open-source python library allows you to integrate Kiln datasets into your own workflows, build fine tunes, use Kiln in Notebooks, build custom tools, and much more! Read the docs for examples.
pip install kiln-aiProducts don’t naturally have “datasets”, but Kiln helps you create one.
Every time you use Kiln, we capture the inputs, outputs, human ratings, feedback, and repairs needed to build high quality models for use in your product. The more you use it, the more data you have.
Your model quality improves automatically as the dataset grows, by giving the models more examples of quality content (and mistakes).
If your product goals shift or new bugs are found (as is almost always the case), you can easily iterate the dataset to address issues.
When building AI products, there’s usually a subject matter expert who knows the problem you are trying to solve, and a different technical team assigned to build the model. Kiln bridges that gap as a collaboration tool.
Subject matter experts can use our easy to use desktop apps to generate structured datasets and ratings, without coding or using technical tools. No command line or GPU required.
Data-scientists can consume the dataset created by subject matter experts, using the UI, or deep dive with our python library.
QA and PM can easily identify issues sooner and help generate the dataset content needed to fix the issue at the model layer.
The dataset file format is designed to be be used with Git for powerful collaboration and attribution. Many people can contribute in parallel; collisions are avoided using UUIDs, and attribution is captured inside the dataset files. You can even share a dataset on a shared drive, letting completely non-technical team members contribute data and evals without knowing Git.
There are new models and techniques emerging all the time. Kiln makes it easy to try a variety of approaches, and compare them in a few clicks, without writing code. These can result in higher quality, or improved performance (smaller/cheaper/faster models at the same quality).
Our current beta supports:
- Various prompting techniques: basic, few-shot, multi-shot, repair & feedback
- Many models: GPT, Llama, Claude, Gemini, Mistral, Gemma, Phi
- Chain of thought prompting, with optional custom “thinking” instructions
- Fine Tuning: create custom models using your Kiln dataset
In the future, we plan to add more powerful no-code options like evals, and RAG. For experienced data-scientists, you can create these techniques today using Kiln datasets and our python library.
See CONTRIBUTING.md for information on how to setup a development environment and contribute to Kiln.
- Python Library: MIT License
- Python REST Server/API: MIT License
- Desktop App: free to download and use under our EULA, and source-available. License
- The Kiln names and logos are trademarks of Chesterfield Laboratories Inc.
Copyright 2024 - Chesterfield Laboratories Inc.