MicroThinker-3B-Preview

MicroThinker-3B-Preview, a new model fine-tuned from the huihui-ai/Llama-3.2-3B-Instruct-abliterated model, focused on advancing AI reasoning capabilities.

Use with ollama

You can use huihui_ai/microthinker directly

ollama run huihui_ai/microthinker:3b

Training Details

This is just a test, but the performance is quite good.

Now, I'll introduce the test environment.

The model was trained using 1 RTX 4090 GPU(24GB) .

The fine-tuning process used 142k from the FineQwQ-142k dataset, max_length(tokens) 21710, quant_bits 4.

The SFT (Supervised Fine-Tuning) process is divided into several steps, and no code needs to be written.

1. Create the environment.

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

git clone https://github.com/modelscope/ms-swift.git

cd ms-swift
pip install -e .
cd ..

2. Download the model and dataset.

huggingface-cli download huihui-ai/Llama-3.2-3B-Instruct-abliterated --local-dir ./huihui-ai/Llama-3.2-3B-Instruct-abliterated
huggingface-cli download --repo-type  dataset huihui-ai/FineQwQ-142k --local-dir ./data/FineQwQ-142k

3. Used only the huihui-ai/FineQwQ-142k, Trained for 1 epoch:

swift sft --model huihui-ai/Llama-3.2-3B-Instruct-abliterated --model_type llama3_2 --train_type lora --dataset "data/FineQwQ-142k/FineQwQ-142k.jsonl" --num_train_epochs 1 --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --max_length 21710 --quant_bits 4 --bnb_4bit_compute_dtype bfloat16 --bnb_4bit_quant_storage bfloat16 --lora_rank 8 --lora_alpha 32 --gradient_checkpointing true --weight_decay 0.1 --learning_rate 1e-4 --gradient_accumulation_steps 16 --eval_steps 100 --save_steps 100 --logging_steps 20 --system "You are a helpful assistant. You should think step-by-step." --output_dir output/MicroThinker-3B-Preview/lora/sft --model_author "huihui-ai" --model_name "MicroThinker-3B-Preview"

4. Save the final fine-tuned model. After you're done, input exit to exit. Replace the directories below with specific ones.

swift infer --model huihui-ai/Llama-3.2-3B-Instruct-abliterated --adapters output/Llama-3.2-3B-Instruct-abliterated/lora/sft/v0-20250106-193759/checkpoint-8786 --stream true --merge_lora true 

This should create a new model directory: checkpoint-8786-merged, Rename the directory to MicroThinker-3B-Preview, Copy or move this directory to the huihui directory.

5. Perform inference on the final fine-tuned model.

swift infer --model huihui/MicroThinker-3B-Preview --stream true --infer_backend pt --max_new_tokens 8192

6. Test examples.

How many 'r' characters are there in the word "strawberry"?