SeongryongJung/Qwen3-4B-Physics-RLSD-TR

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 2, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SeongryongJung/Qwen3-4B-Physics-RLSD-TR is a 4 billion parameter Qwen3-based language model specifically fine-tuned for physics-related tasks using the RLSD_TR method. This model demonstrates specialized performance on the SciKnowEval physics dataset, achieving a validation mean@16 score of 71.09%. It is optimized for scientific reasoning and question-answering within the domain of physics, leveraging a 32768 token context length for complex problem-solving.

Loading preview...

Overview

This model, SeongryongJung/Qwen3-4B-Physics-RLSD-TR, is a specialized variant of the Qwen3-4B architecture, fine-tuned using the RLSD_TR (Reinforcement Learning with Self-Distillation and Trust-Region) method. It is specifically designed and optimized for tasks within the domain of physics, leveraging a 32,768 token context length.

Key Capabilities

  • Physics Domain Expertise: Fine-tuned on the SciKnowEval physics dataset, indicating strong performance in physics-related reasoning and question-answering.
  • Performance: Achieved a best validation mean@16 score of 71.09% on the Physics dataset after 100 training steps.
  • Training Method: Utilizes the RLSD_TR method, which incorporates teacher regularization and a trust-region approach for policy optimization.
  • Efficient Rollout: Employs vLLM for efficient rollouts during training, with a bfloat16 dtype for performance.

Good For

  • Scientific Question Answering: Ideal for applications requiring accurate responses to physics-related queries.
  • Educational Tools: Can be integrated into platforms for learning or assessing physics knowledge.
  • Research Assistance: Useful for researchers needing to process or generate text related to physics concepts and problems.

Finetunes
1 models