| | --- |
| | license: cc-by-4.0 |
| | language: |
| | - en |
| | base_model: Qwen/Qwen2.5-1.5B |
| | pipeline_tag: text-generation |
| | library_name: transformers |
| | tags: |
| | - DeepMiddleGo |
| | - math-reasoning |
| | - fine-tuned |
| | - qwen |
| | model-index: |
| | - name: Mobile-ReasoningLLM-v0-1.5B |
| | results: |
| | - task: |
| | type: text-generation |
| | name: Math Reasoning |
| | dataset: |
| | name: AIME 2024 |
| | type: aime-2024 |
| | metrics: |
| | - name: Pass@1 (avg16) |
| | type: pass@1 |
| | value: 90.0 |
| | - task: |
| | type: text-generation |
| | name: Math Reasoning |
| | dataset: |
| | name: AIME 2025 |
| | type: aime-2025 |
| | metrics: |
| | - name: Pass@1 (avg16) |
| | type: pass@1 |
| | value: 76.7 |
| | --- |
| | # Mobile-Flash-v1-1.5B |
| |
|
| | ## Model Description |
| | Mobile-Flash-v1-1.5B is a fine-tuned derivative of [Qwen2.5-1.5B](https://huggingface.co/Qwen/Qwen2.5-1.5B), optimized for reasoning tasks in mathematics generation. It supports up to 40K output tokens for math problems. This model is designed for both commercial and non-commercial research use. |
| | This repository contains the evluation code of Mobile-Flash-v1-1.5B which **starts to explore Self-RL learning** besides of **sparse reward** learning in the reinforcement learning. In this work, I start to explore the self-rl training algorithm after pre-training, r1-reinforcement learning, |
| | r1-curriculumn reinforcement learning to reduce the difficulty of sparse reward and inefficiency in the RL-Post training stage. |
| |
|
| | - **Architecture**: Dense decoder-only Transformer |
| | - **Base Model**: Qwen2.5-1.5B |
| | - **Parameters**: 1.5 billion |
| | - **Version**: v1 (released Feb 12, 2026) |
| |
|
| | ## Intended Use |
| | - **Primary Use**: Solving complex math problems. |
| | - **Applications**: Research, education, software development, and math reasoning tasks. |
| | - **Limitations**: May not handle ambiguous or poorly formatted inputs well. Ethical use is encouraged to avoid harmful applications. |
| |
|
| | ## Benchmarks |
| | The model was post-trained on a hybrid dataset (automated, human, synthetic) including: |
| | - Math datasets: AIME 2024, AIME 2025 |
| |
|
| | ## Evaluation |
| | The model was evaluated on the following benchmarks, achieving strong performance pass1@avg16: |
| |
|
| | | Model(1.5B) | AIME24 | AIME25 | |
| | |--------------------------|--------|--------| |
| | | Mobile-ReasoningLLM-v0-1.5B | 60.0 | 45.0 | |
| | | Mobile-Flash-ReasoningLLM-v0-1.5B | 70.0 | 60.0 | |
| | | Viber-Thinker-1.5B | 78.0 | 70.0 | |
| | | **Mobile-Flash-v1-1.5B** | **90.0** | **76.7** | |
| | | Model(>235B) | AIME24 | AIME25 | |
| | | GPT-5.2 |97.0+ |97.0+ | |
| | | Grok-4 |97.0+ |97.0+ | |
| | | Gemini-3-Pro |97.0+ |97.0+ | |
| | | GPT-OSS-120B | 96.6 | 97.9 | |
| | | GPT-OSS-20B | 96.0 | 98.7 | |
| | | Grok 3 Mini | 95.8 | 93.3 | |
| | | o4-mini | 93.4 | 92.7 | |
| | | o3 | 91.6 | 86.5 | |
| | | DeepSeek-R1-0528(671B) | 91.4 | 87.5 | |
| | | Qwen-3(235B) | 85.7 | 81.5 | |
| |
|
| | ## How to Use |
| | ### Requirements |
| | - **Library**: `transformers`, `torch`, `vLLM` or `TensorRT-LLM` |
| | - **Hardware**: Trained and Tested on NVIDIA 8xA100-80GB GPUs |
| | - **Environment**: Python 3.10+ (e.g., Conda `hug` environment) |
| |
|
| | ### Inference Example |
| | ```python |
| | import transformers |
| | import torch |
| | model_id = "deepgo/Mobile-Flash-v1-1.5B" |
| | pipeline = transformers.pipeline( |
| | "text-generation", |
| | model=model_id, |
| | model_kwargs={"torch_dtype": torch.bfloat16}, |
| | device_map="auto", |
| | ) |
| | # Math problem prompt |
| | prompt = """Solve the following math problem. Make sure to put the answer (and only answer) inside \\boxed{}.""" |
| | max-length=40,000 is recommend.(reduced from 48,000 to 40,000) |
