README.md

---
datasets:
- PowerInfer/QWQ-LONGCOT-500K
- PowerInfer/LONGCOT-Refine-500K
base_model:
- Qwen/Qwen2.5-3B-Instruct
pipeline_tag: text-generation
language:
- en
library_name: transformers
---
# SmallThinker-3B-preview

We introduce **SmallThinker-3B-preview**, a new model fine-tuned from the [Qwen2.5-3b-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) model. 

## Benchmark Performance

| Model | AIME24 | AMC23 | GAOKAO2024_I | GAOKAO2024_II | MMLU_STEM | AMPS_Hard | math_comp |
|---------|--------|-------|--------------|---------------|-----------|-----------|-----------|
| Qwen2.5-3B-Instruct | 6.67 | 45 | 50 | 35.8 | 59.8 | - | - |
| SmallThinker | 16.667 | 57.5 | 64.2 | 57.1 | 68.2 | 70 | 46.8 |
| GPT-4o | 9.3 | - | - | - | 64.2 | 57 | 50 |

Limitation: Due to SmallThinker's current limitations in instruction following, for math_comp we adopt a more lenient evaluation method where only correct answers are required, without constraining responses to follow the specified AAAAA format.

Colab Link: [Colab](https://colab.research.google.com/drive/182q600at0sVw7uX0SXFp6bQI7pyjWXQ2?usp=sharing)
## Intended Use Cases

SmallThinker is designed for the following use cases:

1.  **Edge Deployment:** Its small size makes it ideal for deployment on resource-constrained devices.
2.  **Draft Model for QwQ-32B-Preview:** SmallThinker can serve as a fast and efficient draft model for the larger QwQ-32B-Preview model. From my test, in llama.cpp we can get 70% speedup (from 40 tokens/s to 70 tokens/s).

## Training Details

The model was trained using 8 H100 GPUs with a global batch size of 16. The specific configuration is as follows:

The SFT (Supervised Fine-Tuning) process was conducted in two phases:

1. First Phase:
   - Used only the PowerInfer/QWQ-LONGCOT-500K dataset
   - Trained for 1.5 epochs
```
### model
model_name_or_path: /home/syx/Qwen2.5-3B-Instruct

### method
stage: sft
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z3_config.json

### dataset
dataset: o1-v2
template: qwen
neat_packing: true
cutoff_len: 16384
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: saves/qwen2-01-qat/full/sft
logging_steps: 1
save_steps: 1000
plot_loss: true
overwrite_output_dir: true
```
2. Second Phase:
   - Combined training with PowerInfer/QWQ-LONGCOT-500K and PowerInfer/LONGCOT-Refine datasets
   - Continued training for 2 additional epochs
```
### model
model_name_or_path: saves/qwen2-01-qat/full/sft/checkpoint-24000

### method
stage: sft
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z3_config.json

### dataset
dataset: o1-v2, o1-v3
template: qwen
neat_packing: true
cutoff_len: 16384
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: saves/qwen2-01-qat/full/sft
logging_steps: 1
save_steps: 1000
plot_loss: true
overwrite_output_dir: true
```

## Limitations & Disclaimer

Please be aware of the following limitations:

*   **Language Limitation:** The model has only been trained on English-language datasets, hence its capabilities in other languages are still lacking.
*   **Limited Knowledge:** Due to limited SFT data and the model's relatively small scale, its reasoning capabilities are constrained by its knowledge base.
*   **Unpredictable Outputs:** The model may produce unexpected outputs due to its size and probabilistic generation paradigm. Users should exercise caution and validate the model's responses.
*   **Repetition Issue:** The model tends to repeat itself when answering high-difficulty questions. Please increase the `repetition_penalty` to mitigate this issue.