83 lines
3.4 KiB
Markdown
83 lines
3.4 KiB
Markdown
---
|
||
license: other
|
||
license_name: qwen-research
|
||
license_link: LICENSE
|
||
base_model: Qwen/Qwen2.5-3B
|
||
library_name: transformers
|
||
pipeline_tag: text-generation
|
||
language:
|
||
- en
|
||
datasets:
|
||
- XINLI1997/WirelessMATHBench-XL
|
||
tags:
|
||
- wireless-communications
|
||
- mathematical-reasoning
|
||
- grpo
|
||
- reinforcement-learning
|
||
- neurips-2026
|
||
---
|
||
|
||
# WirelessMathLM-3B
|
||
|
||
GRPO-trained reference checkpoint from **WirelessMathBench-XL: An Auditable Benchmark for Wireless Mathematical Reasoning** (NeurIPS 2026, Evaluations and Datasets Track).
|
||
|
||
[Project page](https://lixin.ai/WirelessMathBench-XL/) · [arXiv](https://arxiv.org/abs/2509.23219) · [OpenReview](https://openreview.net/forum?id=KbbNUVldqT) · [Code](https://github.com/LiXin97/WirelessMathBench-XL) · [Dataset](https://huggingface.co/datasets/XINLI1997/WirelessMATHBench-XL)
|
||
|
||
## Model
|
||
|
||
- **Base:** [Qwen/Qwen2.5-3B](https://huggingface.co/Qwen/Qwen2.5-3B). Initialised from the **Qwen2.5-3B** base checkpoint (no SFT warm-start).
|
||
- **Training:** GRPO with [EasyR1](https://github.com/hiyouga/EasyR1) on the WirelessMathBench-XL train split (3,227 problems), 40 epochs / 240 steps, composite reward 0.1 × format + 0.9 × verifier accuracy, AdamW (lr 1e-6, cosine), KL coefficient 0.01, 4 × NVIDIA RTX 4090 D. See Appendix B of the paper.
|
||
- **Precision:** bfloat16.
|
||
|
||
## Results
|
||
|
||
| Evaluation (WirelessMathBench-XL test) | Accuracy |
|
||
|---|---:|
|
||
| Full 800-item test split (paper Tab. 3, row D13) | 25.50% |
|
||
| 310-item public test subset (paper Appendix L) | 25.16% |
|
||
|
||
Locked protocol: 2k-token answer budget, T = 0.6, GPT-4.1-mini judge on every response.
|
||
|
||
Training-time check at the final checkpoint (step 240; paper Appendix R): base 12.37% → GRPO 25.62%. These values use a different protocol from the locked scores.
|
||
|
||
## Usage
|
||
|
||
The paper queries these checkpoints through a raw completion endpoint (no chat template), using the dataset's `prompt` field.
|
||
|
||
```python
|
||
from datasets import load_dataset
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
repo = "XINLI1997/WirelessMathLM-3B"
|
||
tok = AutoTokenizer.from_pretrained(repo)
|
||
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
|
||
|
||
ex = load_dataset("XINLI1997/WirelessMATHBench-XL", "cc_by", split="test")[0]
|
||
prompt = ex["prompt"]
|
||
inputs = tok(prompt, return_tensors="pt").to(model.device)
|
||
out = model.generate(**inputs, max_new_tokens=2048, do_sample=True, temperature=0.6)
|
||
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
|
||
```
|
||
|
||
## Limitations
|
||
|
||
- These are **release-split learnability checks**, not evidence of general or verifier-independent reasoning. The train/test split is problem-level, so train and test share source papers.
|
||
- The reward verifier and the evaluation fallback share the same checkpoint family and prompt style.
|
||
- Intended for research on wireless mathematical reasoning; not a general-purpose assistant.
|
||
|
||
## License
|
||
|
||
This model is derived from Qwen2.5-3B and is released under the [Qwen Research License](LICENSE) inherited from the base model.
|
||
|
||
## Citation
|
||
|
||
```bibtex
|
||
@inproceedings{
|
||
li2026wirelessmathbenchxl,
|
||
title={WirelessMathBench-XL: An Auditable Benchmark for Wireless Mathematical Reasoning},
|
||
author={Xin Li and Mengbing Liu and Yiyang Zhu and Wenhe Zhang and Li Wei and Jiancheng An and Chau Yuen},
|
||
booktitle={The Fortieth Annual Conference on Neural Information Processing Systems Evaluations and Datasets Track},
|
||
year={2026}
|
||
}
|
||
```
|