SmallThinker-3B-preview

We introduce SmallThinker-3B-preview, a new model fine-tuned from the 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.

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:

neat_packing: true
cutoff_len: 16384
per_device_train_batch_size: 2
gradient_accumulation_steps: 1
learning_rate: 1.0e-5
num_train_epochs: 3
lr_scheduler_type: cosine
warmup_ratio: 0.02
bf16: true
ddp_timeout: 180000000
weight_decay: 0.0

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

2. Second Phase:

   • Combined training with PowerInfer/QWQ-LONGCOT-500K and PowerInfer/LONGCOT-Refine datasets
   • Continued training for an additional 2 epochs

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.