localtrainer: No-Code AI Trainer

This repository hosts a comprehensive Gradio-based application that allows you to:

• Chat with an AI model (supporting both Simple and "Teaching" mode)
• Perform web searches to inject external knowledge into your prompts
• Generate multi-step Chain-of-Thought (CoT) reasoning
• Collect and store user feedback (including step-level ratings for CoT)
• Conduct advanced training modes (RLHF, Supervised Fine Tuning, Teacher Fine Tuning)
• Manage RAG (Retrieval-Augmented Generation) instructions
• Maintain a feedback history and a training dashboard

Features

1. Model Chat

   • Simple Conversation mode provides concise answers.
   • Teaching mode uses multi-step reasoning, optionally including parallel CoTs and web search.

2. Parallel or Sequential CoTs

   • If system memory/VRAM allows, generate two Chains-of-Thought in parallel (one from a teacher model if configured).

3. Feedback Storage

   • All feedback and CoT step-level ratings can be saved for later analysis or fine-tuning.

4. Training Monitor

   • Supports advanced fine-tuning modes: RLHF, supervised fine-tuning, or teacher-based fine-tuning.

5. Prompt Engineering

   • Combine arbitrary prompts with RAG instructions, then run them through any loaded model.

6. RAG Management

   • Easily create, load, and update RAG instructions stored in rag_data.json.

7. Feedback History

   • View, edit, and delete feedback entries stored in feedback_dataset.json.

Installation

1. Clone the Repository

   git clone https://github.com/stacker00/localtrainer.git
   cd localtrainer

2. Install Dependencies

   • Create or activate a virtual environment (recommended), then install the required packages:

   pip install -r requirements.txt

Usage

1. Run the Application

   python main.py

   • This will launch a Gradio interface at a local URL (e.g., http://127.0.0.1:7860).

2. Tabs Overview

• Chat with Model: Enter text, select your mode (Simple Conversation / Teaching), optionally include a web search, choose a model, and click Submit.
• CoT Splitting & Ratings: Fetch the generated CoTs from the chat tab to split and rate them step-by-step, then save.
• Training Monitor: Perform advanced training steps (RLHF, Supervised, Teacher).
• Prompt Engineering: Combine custom prompts with optional RAG instructions, then generate outputs.
• RAG Instructions: Manage your RAG instructions stored in rag_data.json.
• Dashboard: Quickly view memory usage and feedback statistics.
• Feedback History: Edit or delete prior feedback rows stored in feedback_dataset.json.

Testing

   pytest test_main.py

or

   python -m unittest test_main.py

Issues

• Web search sometimes works when duckduckgo search page is opened on a browser. good luck.
• Metamask not working yet as of first commit.
• It is full of other bugs since I just wrote it. Hope to improve them soon.