Skip to content

README.md

Latest commit

 

History

History
76 lines (46 loc) · 2.1 KB

README.md

File metadata and controls

76 lines (46 loc) · 2.1 KB

Atom of Thoughts (AoT) Fine-Tuning System

Industrial-grade fine-tuning system implementing the Atom of Thoughts (AoT) reasoning framework. Designed to run on limited-resource devices using PEFT and 4-bit quantization.

What this project does

Standard Chain-of-Thought (CoT) prompting is linear and memory-heavy. AoT treats reasoning as a Markovian process:

  1. Decomposition — breaks a complex problem into independent atomic states
  2. Atomic Reasoning — solves each state in isolation, preventing history interference
  3. Contraction — merges atomic solutions into a final, verifiable answer

This reduces token bloat and improves reasoning accuracy on complex tasks.

System architecture

  • src/training/aot_engine.py — Core Decompose-Solve-Contract loop
  • src/model/model_manager.py — Model loading with MXFP4 and NF4 4-bit quantization
  • src/training/trainer.py — Unified training pipeline for multi-scale models (0.5B to 20B)
  • local_lite/ — Optimized sub-system for CPU-only and 16GB RAM environments

Getting started

Prerequisites

pip install -r requirements.txt

Training with AoT

python src/training/trainer.py --model qwen_7b --template aot --dataset ./data/aot_research_data.json

Set WANDB_MODE=disabled if you are not using Weights and Biases:

WANDB_MODE=disabled python src/training/trainer.py --model qwen_7b --template aot --dataset ./data/aot_research_data.json

Sync changes to Git

To push local changes to the repository, run the sync script from the project root:

./scripts/sync_to_git.ps1

This script stages all modified files, commits with a timestamp message, and pushes to the main branch.

Current status

  • Smoke tests confirmed successful inference on Qwen2.5-0.5B with 4-bit NF4 quantization on CPU
  • Training loop runs end-to-end with AoT decompose-solve-contract prompt structure
  • Full evaluation on a held-out reasoning benchmark is in progress

Tech stack

Python · PyTorch · Transformers · PEFT · BitsAndBytes · Qwen2.5 · WandB

License

MIT