SeongryongJung/qwen3-4b-material-grpo

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

The SeongryongJung/qwen3-4b-material-grpo model is a 4 billion parameter language model fine-tuned from Qwen/Qwen3-4B. It utilizes the GRPO method on the 'material' split of the sciknoweval dataset, achieving a peak validation performance of 78.32% mean@16. This model is specialized for tasks related to material science, demonstrating optimized performance on scientific knowledge evaluation benchmarks.

Loading preview...

Model Overview

SeongryongJung/qwen3-4b-material-grpo is a 4 billion parameter language model, fine-tuned from the base Qwen/Qwen3-4B architecture. This model has undergone specialized training using the GRPO (Generalized Reinforcement Learning from Policy Optimization) method, specifically on the material split of the sciknoweval dataset.

Key Characteristics

  • Base Model: Qwen3-4B, a robust foundation for language understanding.
  • Fine-tuning Method: GRPO, indicating a reinforcement learning approach to optimize performance.
  • Specialized Domain: Trained on the material split of sciknoweval, suggesting expertise in material science-related knowledge and evaluation tasks.

Performance Highlights

Validation performance was tracked using the val-aux/sciknoweval/reward/mean@16 metric. The model achieved a best mean@16 score of 78.32% at step 40 during its training process. The final uploaded weights correspond to global_step_100/actor, which recorded a mean@16 of 66.16%.

Intended Use Cases

This model is particularly suited for applications requiring knowledge and reasoning within the domain of material science. Its fine-tuning on the sciknoweval dataset's material split implies strong capabilities in tasks such as:

  • Scientific knowledge evaluation in material science.
  • Answering questions related to material properties, synthesis, and applications.
  • Assisting with research and analysis in the materials domain.