LIMA-GUI

LIMA-GUI is a simple utility for gathering LIMA-like data (of extremely high quality) to train you own personal AI (based on LLAMA, Mistral or others).

It uses a simple data format that's fully compliant with OpenAI API and natively supported by Huggingface trainers. Chat is represented as a list of JSONs:

[
    {
        "role": "role",
        "content": "content"
    }
]

Example chat:

[
    {
        "role": "user",
        "content": "Hey, what is 4 + 4?"
    },
    {
        "role": "assistant",
        "content": "Hey! 4 + 4 is 8."
    }
]

Usage

1. Download latest release (main branch is likely to be less stable).
2. Install lima-gui by running pip install . in the repo folder.
3. Run limagui.

If you experience any problems, please make a corresponding issue.

Motivation

I strongly believe in personalized AI tailored to the user's personal preferences and use cases. Having your own AI is way better instead of relying on others' work as you have full control and customization freedom. Though making one takes considerable effort and lots of compute. It used to take tens of thousands of examples (at the very least) to train an AI assistant, but now we have:

• LIMA: Less Is More for Alignment. The paper has shown that it is possible to create a decent AI assistant using only a handful of curated examples (around 1000 samples). kaiokendev confirmed that LIMA actually works kaiokendev's findings.
• LORA: LOW-RANK ADAPTATION OF LARGE LANGUAGE MODELS and QLORA: Efficient Finetuning of Quantized LLMs. No more bazilion gigabytes of memory needed, a high-grade consumer level GPU (like RTX 3090) should do the trick.

So if the off-the-shelf AIs aren't doing it for you, here is the recipe:

1. Good foundational LLM (which are plenty right now).
2. Some compute (even colab might suffice).
3. Your own unique data.

Now, the last part is a bit tricky. In my case I need a hand-crafted multi-turn chat dataset (public ones don't do). From what I've seen, there's a real lack of tools to help collect this kind of data. A lot of them are stuck in the basic Alpaca format or can be a bit unwieldy—great for big teams, maybe not so much for solo devs.

That's why I decided to make my own tool, keeping things straightforward:

• A user-friendly UI for managing chat data.
• An easy-to-use data storage format that's on good terms with huggingface.
• Integration with the OpenAI API to help with data collection.

LIMA-GUI functionality

Note

Current UI visualizations are bit outdated as I added support for function calling.

LIMA-GUI allows you to manage manually curated (basically hand written) chat data. Main window:

I guess the UI here is mostly self-explanatory, though here is an explicit list of things you can see/do:

• In the main window there is a list of chats.
• You can add, delete, copy chats.
• In the file menu you'll see Open, Save current, Save to and Export as OpenAI dataset options. The former 3 are self-explanatory. Export as OpenAI dataset exports your dataset in OpenAI finetuning API compliant format (jsonl with a lot of `{"messages": [...]}). This way you can directly feed the exported data in the finetuning API or use it with huggingface trainer (I believe it supports this format by default).
• In the settings menu you can tweak some of the app settings as well as OpenAI API configuration.

Double-clicking on a chat opens a chat window:

Chat is represented as a list of messages and can be seen on the left. Each message is given a role:

• system - system prompt (for steering the model's behaviour, priming, etc).
• user - user message.
• assistant - AI's reply.

On the right you can see some of the chat's properties:

• Name - a short description of the chat.
• Language - primary language in the chat (currently ru and en are supported, I'll make it configurable later).
• Tokens - approximate number of tokens in the chat. It is being computed using a huggingface tokenizer that's set in the settings menu ('Mistral-7b' tokenizer is used by default). I need my AI to operate with long contexts, so this stat is really helpful for me.
• Msg count - number of messages in the chat.
• Tags - well, it's a list of tags for the chat. It helps to classify individual chats and split the dataset appropriately if needed.

The coolest thing is that while crafting your own data, you can use OpenAI's models (or any other that are hosted via OpenAI API) to assist you during the process. You have two options:

1. Generate full assistant's reply (using chat API).
2. Generate a part of the reply (using completion API).

Other important things:

• Shortcut for saving current dataset: Ctrl + S.
• Shortcut for saving current dataset in a new file: Ctrl + Shift + S.
• Shortcut for opening a dataset file: Ctrl + R.
• You can replace the API base for the OpenAI API. Meaning you can host you own model to help you in your endeavors (I haven't tested it yet though, but it should work). The only requirement is that the hosting framework of your choice should mimic OpenAI API format. I am using SimpleAI/vllm and it works great for me.
• You can open datasets that are OpenAI finetuning API compliant (jsonl with a lot of `{"messages": [...]}). All chats will have the same titles (I plan on adding some algo to generate meaningful titles) and other metadata will be absent, but you can add it if needed. Other than that, you can work with the data as usual.
• When using completions API, LIMA-GUI formats the chat data in ChatML format. Beware of that when using custom models.

Function calling / Tools

Warning

This functionality is experimental at the moment and you may encounter data corrupting bugs. Make sure you have a running copy of your data before using this functionality.

LIMA-GUI also supports function calling API:

• You can define multiple functions within a chat (python code executors, data retrieval systems, etc.).
• Assistant messages may contain function calls.
• Function calling is integrated with OpenAI API so the calls may be generated automatically.

Note

LIMA-GUI does not execute the functions, it only holds their corresponsing representations. You'll have to manually execute requests and pass into the chat their execution results.

Data format:

• functions defined inside a conversation as:

{
    "type": "function",
    "function": {
        "name": "my_function",
        "description": "My function",
        "parameters": {
            "type": "object",
            "properties": {
                "param1": {
                    "type": "string",
                    "description": "First parameter"
                },
                "param2": {
                    "type": "string",
                    "enum": ["value1", "value2"]
                }
            },
            "required": ["param1"]
        }
    }
}

Note

description and enum fields can be skipped.

Note

Functions' underlying data format is (supposedly) fully compliant with OpenAI API and can be used as is.

• messages

[
    {
        "role": "assistant",
        "content": "content (cant be none)",
        "function_call": {
            "name": "function_name",
            "arguments": {
                "arg1_name": "value",
                "arg2_name": "value"
            }
        }
    },
    {
        "role": "function",
        "name": "function_name",
        "content": "function's response"
    }
]

Note

Messages' underlying data format is fully compliant with OpenAI API and can be used as is.

See examples folder for a sample of chats with function calling.

Caution

Note

This section used to be longer!

The project is still in its early stages, so expect a few bumps along the road. Despite how yong the project is, I've found current feature set to be super handy (at least for my needs).

TODO

• QoL
  • Focus main window when closing a chat window.
  • Fix coloring, it's too contrast and sometimes makes things difficult to see.
  • Ctrl + S should work in the chat window as well.
  • Reduce the size of the right side in chat.
• Safeguards not to lose data.
  • Automatic saving (optional).
  • Asking to save if haven't already when closing app.
  • Asking to save if opening a new file.
  • Ctrl + S should save in an already opened file.
  • Ctrl + Shift + S should save in a new file.
  • Show somewhere a little message "Saved".
• Somehow refactor current shitty chat UI.
  • Make scrolling smooth.
  • Make list items fit message contents.
  • Scrolling at the end of the list on opening.
  • Select/Focus newly added item.
  • Scroll down the focused ChatItem during AI text generation.
  • Automatically focus the item you are typing in.
• Huggingface integration (download/upload the dataset).
• Deletion of multiple messages.
• Token count using Huggingface tokenizers (Mistral tokenizer by default).
• Manual on how to use lima-gui (!).
• Tags for chats (like coding, logic, qa, etc).
• Default dataset config (contains config for languages, tags and tokenizer).
• OpenAI API integration to allow for AI generated answers.
  • Chat API support.
  • Completion API support.
  • Token count support.
  • Stopping generation (!).
  • Function calling.
    • Proper UI for function call messages.
    • Function roles.
    • Data integrity when changing functions on the fly (!).
• Keep tabulation.
• Save settings/current config.
• Stats board that shows various useful statistics about current data (token count distribution, number of samples by tags/languages).
• Remove stupid prints and add proper logging.
• Explain tags.
• Save datasets in OpenAI format ('.jsonl').
• Load datasets in OpenAI format ('.jsonl').

Pull requests

Any changes that simplify or make data gathering more comfortable (UI/UX, QoL, tools, etc) are welcome. Refactoring suggestions that simplfy/decouple/better structure the code are also welcome.