MicroThinker-1B-Preview

MicroThinker-1B-Preview, a new model fine-tuned from the huihui-ai/Llama-3.2-1B-Instruct-abliterated model, focused on advancing AI reasoning capabilities.

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 only 20,000 records from each dataset.

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-3.2-1B-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. Used only the huihui-ai/QWQ-LONGCOT-500K dataset (#20000), Trained for 1 epoch:

swift sft --model huihui-ai/Llama-3.2-1B-Instruct-abliterated --model_type llama3_2 --train_type lora --dataset "data/QWQ-LONGCOT-500K/qwq_500k.jsonl#20000" --torch_dtype bfloat16 --num_train_epochs 1 --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --learning_rate 1e-4 --lora_rank 8 --lora_alpha 32 --target_modules all-linear --gradient_accumulation_steps 16 --eval_steps 50 --save_steps 50 --save_total_limit 2 --logging_steps 5  --max_length 16384  --output_dir output/Llama-3.2-1B-Instruct-abliterated/lora/sft --system "You are a helpful assistant. You should think step-by-step." --warmup_ratio 0.05 --dataloader_num_workers 4 --model_author "huihui-ai" --model_name "MicroThinker"

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

swift infer --model huihui-ai/Llama-3.2-1B-Instruct-abliterated --adapters output/Llama-3.2-1B-Instruct-abliterated/lora/sft/v0-20250102-153619/checkpoint-1237 --stream true --merge_lora true 

This should create a new model directory: checkpoint-1237-merged, Copy or move this directory to the huihui directory.

5. Perform inference on the fine-tuned model.

swift infer --model huihui/checkpoint-1237-merged --stream true --infer_backend pt --max_new_tokens 8192

6. Combined training with huihui-ai/QWQ-LONGCOT-500K (#20000) and huihui-ai/LONGCOT-Refine datasets (#20000), Trained for 1 epoch:

swift sft --model huihui-ai/checkpoint-1237-merged --model_type llama3_2 --train_type lora --dataset "data/QWQ-LONGCOT-500K/qwq_500k.jsonl#20000" "data/LONGCOT-Refine-500K/refine_from_qwen2_5.jsonl#20000" --torch_dtype bfloat16 --num_train_epochs 1 --per_device_train_batch_size 1 --per_device_eval_batch_size 1 --learning_rate 1e-4 --lora_rank 8 --lora_alpha 32 --target_modules all-linear --gradient_accumulation_steps 16 --eval_steps 50 --save_steps 50 --save_total_limit 2 --logging_steps 5  --max_length 16384  --output_dir output/Llama-3.2-1B-Instruct-abliterated/lora/sft2 --system "You are a helpful assistant. You should think step-by-step." --warmup_ratio 0.05 --dataloader_num_workers 4 --model_author "huihui-ai" --model_name "MicroThinker"

7. 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/checkpoint-1237-merged --adapters output/Llama-3.2-1B-Instruct-abliterated/lora/sft2/v0-20250103-121319/checkpoint-2474 --stream true --merge_lora true 

This should create a new model directory: checkpoint-2474-merged, Rename the directory to MicroThinker-1B-Preview, Copy or move this directory to the huihui directory.

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

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

9. Test examples.

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