huyhoangvbck/Qwen3.8-27B-RL-Social-Science
The huyhoangvbck/Qwen3.8-27B-RL-Social-Science model is a 27 billion parameter Qwen/Qwen3.8 base model fine-tuned using Reinforcement Learning with Group Relative Policy Optimization (GRPO). It is specifically optimized for social science tasks, demonstrating enhanced performance on the VMLU Social Science benchmark covering economics, geography, civics, law, and philosophy. This model integrates LoRA weights directly into the base model, making it ready for direct deployment in applications requiring specialized social science reasoning.
Loading preview...
Model Overview
The huyhoangvbck/Qwen3.8-27B-RL-Social-Science is a specialized large language model built upon the Qwen/Qwen3.8-27B base architecture. It has undergone Reinforcement Learning (RL) using the Group Relative Policy Optimization (GRPO) algorithm, implemented via the Megatron-SWIFT framework.
Key Capabilities & Optimization
- Social Science Specialization: The model is specifically fine-tuned and optimized for tasks within the social sciences, as evidenced by its training on the VMLU Social Science benchmark. This includes domains such as economics, geography, civics, law, and philosophy.
- Reinforcement Learning: Utilizes GRPO for enhanced performance, focusing on improving accuracy and formatting for specific social science question types.
- Integrated LoRA Weights: The LoRA adapter weights (Rank 32, Alpha 32) have been fully merged into the base model, simplifying deployment and usage.
- Training Configuration: Employed a rule-based accuracy reward (
vmlu_acc) combined with a format reward (vmlu_format) during RL training, ensuring both correctness and appropriate output structure.
Recommended Usage
This model is particularly well-suited for applications requiring nuanced understanding and generation within social science contexts. It can be efficiently served using vLLM for high-throughput inference or integrated with Hugging Face Transformers for standard deployments. The model supports a context length of 32768 tokens, making it capable of handling detailed social science queries.