XINLI1997/WirelessMathLM-Qwen3-4B

WirelessMathLM-Qwen3-4B Overview

WirelessMathLM-Qwen3-4B is a specialized 4 billion parameter language model developed by XINLI1997, derived from the Qwen3-4B architecture. It was trained using the GRPO (Generative Reinforcement Policy Optimization) method on the WirelessMathBench-XL dataset, a benchmark focused on auditable wireless mathematical reasoning.

Key Capabilities and Training

• Specialized Reasoning: This model is specifically fine-tuned for mathematical reasoning problems within the wireless communications domain, as evidenced by its training on the WirelessMathBench-XL dataset.
• GRPO Training: The model underwent 40 epochs (240 steps) of GRPO training, utilizing a composite reward function (0.1 × format + 0.9 × verifier accuracy) and a KL coefficient of 0.01.
• Performance Improvement: During training, the GRPO-trained model showed an improvement from a base accuracy of 13.63% to 17.87% on internal checks, indicating enhanced reasoning capabilities in its target domain.
• Precision: The model operates in bfloat16 precision.

Intended Use and Limitations

• Research Focus: This model is primarily intended for research purposes related to wireless mathematical reasoning and serves as a learnability check for the WirelessMathBench-XL benchmark.
• Not General-Purpose: It is not designed as a general-purpose assistant and its reasoning capabilities are specific to the wireless domain problems it was trained on.
• Contextual Dependency: The training and evaluation protocols highlight that its performance is tied to the specific problem-level split and shared source papers within the benchmark, rather than demonstrating broad, verifier-independent reasoning.