Model description
This model is a fine-tuned version of Qwen/Qwen2.5-7B-Instruct on the Bespoke-Stratos-17k dataset. The dataset is derived by distilling DeepSeek-R1 using the data pipeline of Berkeley NovaSky’s Sky-T1 with some modifications. More info in the dataset card at Bespoke-Stratos-17k. It outperforms Qwen-2.5-7B-Instruct on math reasoning benchmarks:
| Bespoke-Stratos-7B | Qwen2.5-7B-Instruct | DeepSeek-R1-Distill-Qwen-7B (Ours) | DeepSeek-R1-Distill-Qwen-7B (Reported) | |
|---|---|---|---|---|
| AIME2024 | 20.0 | 10.0 | 43.3 | 55.5 |
| MATH500 | 82.0 | 74.2 | 89.4 | 92.8 |
| GPQA-Diamond | 37.8 | 33.3 | 44.9 | 49.1 |
| LiveCodeBench v2 Easy | 71.4 | 65.9 | 81.3 | - |
| LiveCodeBench v2 Medium | 25.5 | 18.9 | 42.2 | - |
| LiveCodeBench v2 Hard | 1.6 | 3.3 | 2.4 | - |
| LiveCodeBench v2 All | 36.1 | 31.9 | 46.6 | - |
Note that the authors of Sky-T1 had noted that they saw little or no improvement in training 7B or 14B models with their data. However, see an improvement, though not at the scale of DeepSeek's distilled model. The reason could be that we used 17k examples, while DeepSeek seems to have used 800k.
Intended uses & limitations
Apache 2.0 License
Training procedure
We used 8xH100 to train the model for 7 hours.
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 1
- eval_batch_size: 8
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- gradient_accumulation_steps: 12
- total_train_batch_size: 96
- total_eval_batch_size: 64
- optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_ratio: 0.1
- num_epochs: 3.0
Training results
Framework versions
- Transformers 4.46.1
- Pytorch 2.5.1+cu124
- Datasets 3.1.0
- Tokenizers 0.20.3