⚛️ Atom of Thoughts (AoT) Fine-Tuning System

Industrial-grade fine-tuning system designed to implement and scale the Atom of Thoughts (AoT) reasoning framework on any device, even with limited resources.

🚀 Project Motive

The goal of this project is to democratize high-level reasoning research. By leveraging Parameter-Efficient Fine-Tuning (PEFT) and 4-bit Quantization, we enable researchers to train large models (up to 20B+) on consumer-grade hardware or CPU-only environments.

🧠 What is Atom of Thoughts (AoT)?

Unlike standard Chain-of-Thought (CoT) which is linear and memory-heavy, AoT treats reasoning as a Markovian Process:

1. Decomposition: Breaking complex problems into independent 'atomic' states.
2. Atomic Reasoning: Solving each state in isolation to prevent history-interference.
3. Contraction: Merging atomic solutions into a final, verifiable answer.

This method drastically reduces token bloat and improves reasoning accuracy for complex mathematical and research tasks.

🏗️ System Architecture

• src/training/aot_engine.py: The core orchestrator for the Decompose-Solve-Contract loop.
• src/model/model_manager.py: Handles industrial model loading with native support for MXFP4 and NF4 quantization.
• src/training/trainer.py: Unified training pipeline supporting multi-scale models (0.5B, 7B, 20B).
• local_lite/: Optimized sub-system for ultra-low resource (CPU/16GB RAM) training.

🛠️ Getting Started

Prerequisites

pip install -r requirements.txt

Training AoT (Research Batch)

# Optimized for CPU/Limited Resource
$env:WANDB_MODE="disabled"; python src/training/trainer.py --model qwen_7b --template aot --dataset ./data/aot_research_data.json

🔄 Auto-Sync Protocol

To ensure the repository is always up-to-date, I have implemented an auto-sync utility. You can run it anytime you make changes:

./scripts/sync_to_git.ps1

Note: As your AI assistant, I will automatically run this sync after major upgrades.

📈 Current Status

• ✅ Infrastructure Verified: Smoke tests successful on Qwen2.5-0.5B.
• ✅ Quantization Hardened: Native support for pre-quantized 20B models confirmed.
• ✅ Auto-Sync Active: GitHub repository is linked and automated.

---

Developed for advanced AI research in reasoning and scaling.