🧠 Fine-Tune Dataset Constructor

This tool is published to GitHub pages by this link

A React-based tool for constructing fine-tuning datasets with list and grid forms, featuring the ability to download and upload data as JSONL files compatible with OpenAI's fine-tuning API. This project leverages the react-declarative library to create dynamic, interactive forms for defining user inputs, preferred outputs, non-preferred outputs, and chat history, complete with tool definitions and calls.

Features

• Dynamic Form Interface: Create and edit dataset entries using a combination of list and grid layouts.
• Input Definition: Specify user input messages with configurable tools and parameters.
• Output Configuration: Define preferred and non-preferred outputs with content and associated tools.
• Tool Management: Add up to five tools per input/output with customizable arguments (name, type, description, enum, required status).
• Tool Name Autocomplete: When selecting tools, the UI provides autocomplete for tool names, pulling from defined input tools to streamline selection.
• Tool Enum Autocomplete: For arguments with enum values, the form offers autocomplete for predefined options, enhancing accuracy and speed.
• History with Tools: Include multi-turn conversational context with tool calls in history messages (currently supporting tool1 per message).
• Data Persistence: Save changes to local storage and retrieve them by ID.
• JSONL Support: Import/export datasets as JSONL files with automatic generation of unique tool_call_ids per OpenAI spec.
• Material-UI Integration: Styled with Material-UI components for a modern, responsive UI.
• Breadcrumbs Navigation: Easy navigation with save, back, and draft actions.
• Validation: Robust checks for tool calls, message order, and history tool consistency.
• Performance Optimized: Uses Maps for O(1) lookups in validation and UI callbacks.

Prerequisites

• Node.js (v16 or higher recommended)
• npm or yarn
• Basic familiarity with React and TypeScript

Installation

1. Clone the Repository:

   git clone https://github.com/<your-username>/<your-repo-name>.git
   cd <your-repo-name>

2. Install Dependencies: Using npm:

   npm install

   Or using yarn:

   yarn install

3. Run the Application:

   npm start

   Or with yarn:

   yarn start

   The app will be available at http://localhost:3000.

Usage

1. Access a Dataset Entry:

   • Create a new entry by using + Create button on a datagrid page
   • Navigate to an entry using datagrid

2. Edit Fields:

   • Chat History: Add up to five messages with roles (user, assistant, system, tool) and content. Assistant messages can include a tool call (tool1).
   • User Input: Enter message content and define up to five tools with parameters.
   • Preferred Output: Specify response content and up to five tool calls with autocomplete for tool names and enum values.
   • Non-Preferred Output: Define an alternative response.

3. Save Changes:

   • Click "Save" in the breadcrumbs to persist to local storage, or "Save as Draft" to store temporarily.
   • Use "Back" to return to the main page, with a confirmation if unsaved changes exist.

4. Export as JSONL:

   • Press three dots on a gridview form and use Export to jsonl

5. Import JSONL:

   • Press three dots on a gridview form and use Import from jsonl

6. Fine-Tune with OpenAI:

   • Export your dataset and upload it:

     openai api fine_tunes.create -t "path/to/your_dataset.jsonl" -m "base_model_name"

Project Structure

• src/OneView.tsx: Main component with form logic and UI.
• src/config/history.ts: Routing configuration.
• src/config/storage.ts: Storage logic for dataset persistence.
• src/config/draft.ts: Draft state management.
• src/validation/: Validation functions (validateToolCalls, validateMessageOrder, validateMessageTools).
• fields: TypedField array defining the form structure.

Dependencies

• @mui/icons-material: Material-UI icons.
• @mui/material: Material-UI components.
• lodash-es: Utility functions (e.g., get for safe property access).
• react-declarative: Declarative form-building library.

Data Structure

• IStorageItem: Fine-tuning example with input, preferred_output, non_preferred_output, and history.
• IToolDefinition: Tool spec with name, description, and up to five arguments.
• ITool: Tool call with name and up to five key-value arguments.
• IHistoryMessage: History message with role, content, and one tool call (tool1).

Technical Highlights

• Tool History: Supports tool calls in history, exported with unique tool_call_ids.
• Tool Name Autocomplete: In createToolOutput, the itemList for tool names dynamically lists available tools from input definitions, enabling quick selection with autocomplete.
• Tool Enum Autocomplete: In createToolOutputArgument, enum-based arguments provide a dropdown with autocomplete, sourced from tool definitions, improving data entry efficiency.

Extending the Project

• Full JSONL UI: Add export/import buttons to <Breadcrumbs2 /> or <One /> components.
• Enhanced Validation: Expand validation rules.
• Backend: Replace local storage with an API.
• Multi-Tool History: Extend IHistoryMessage to support multiple tools.

Contributing

1. Fork the repository.
2. Create a branch (git checkout -b feature/<feature-name>).
3. Commit changes (git commit -m "Add <feature-name>").
4. Push (git push origin feature/<feature-name>).
5. Open a pull request.

License

MIT License - see LICENSE for details.