Micro Thinker 1 B Preview

Introduction

MicroThinker-1B-Preview is a model fine-tuned from the huihui-ai/Llama-3.2-1B-Instruct-abliterated model, aimed at enhancing AI reasoning capabilities. This version is focused on text generation and uses the Transformers library.

Architecture

The model architecture is based on the Llama-3.2 series, specifically the Llama-3.2-1B-Instruct-abliterated variant. It is designed to be uncensored, conversational, and suitable for text-generation tasks, leveraging the advanced features of the Transformers library.

Training

MicroThinker-1B-Preview was trained in a test environment using a single RTX 4090 GPU with 24GB memory. The fine-tuning process employed the SFT (Supervised Fine-Tuning) framework, utilizing 20,000 records from two datasets: QWQ-LONGCOT-500K and LONGCOT-Refine-500K. Training was conducted over one epoch, focusing on specific parameters like learning rate, batch size, and gradient accumulation.

Guide: Running Locally

1. Setup Environment:

   • Create a conda environment:

     conda create -yn ms-swift python=3.11
     conda activate ms-swift
     

   • Clone and install the ms-swift repository:

     git clone https://github.com/modelscope/ms-swift.git
     cd ms-swift
     pip install -e .
     cd ..
     

2. Download Model and Datasets:

   • Use the huggingface-cli to download the model and datasets:

     huggingface-cli download huihui-ai/Llama-3.2-1B-Instruct-abliterated --local-dir ./huihui-ai/Llama-3.2-1B-Instruct-abliterated
     huggingface-cli download --repo-type dataset huihui-ai/QWQ-LONGCOT-500K --local-dir ./data/QWQ-LONGCOT-500K
     huggingface-cli download --repo-type dataset huihui-ai/LONGCOT-Refine-500K --local-dir ./data/LONGCOT-Refine-500K
     

3. Fine-Tuning and Inference:

   • Follow the provided commands for fine-tuning and inference using the swift tool. Consider using cloud GPUs such as AWS EC2 P3 instances or Google Cloud's A100 GPUs for optimal performance.

4. Inference:

   • Perform inference with:

     swift infer --model huihui/MicroThinker-1B-Preview --stream true --infer_backend pt --max_new_tokens 8192
     

License

MicroThinker-1B-Preview is released under the Apache-2.0 License, allowing for open use and modification under its terms.