SeongryongJung/qwen3-4b-biology-rlsd-ema005
SeongryongJung/qwen3-4b-biology-rlsd-ema005 is a 4 billion parameter Qwen3-based language model fine-tuned specifically for biology-related tasks. It utilizes Reinforcement Learning from Scientific Data (RLSD) with an EMA of 0.05 on the biology split of the SciKnowEval dataset. This model demonstrates specialized performance in scientific knowledge evaluation within the biology domain, achieving a peak mean@16 validation score of 53.50%. It is optimized for applications requiring nuanced understanding and generation of biological information.
Loading preview...
Model Overview
SeongryongJung/qwen3-4b-biology-rlsd-ema005 is a specialized 4 billion parameter language model built upon the Qwen/Qwen3-4B architecture. This model has undergone fine-tuning using Reinforcement Learning from Scientific Data (RLSD) with an Exponential Moving Average (EMA) of 0.05, specifically targeting the biology split of a scientific knowledge evaluation dataset.
Key Capabilities & Performance
- Biology Domain Specialization: Fine-tuned to excel in tasks related to biological knowledge and understanding.
- RLSD Training: Leverages Reinforcement Learning from Scientific Data to enhance performance in scientific contexts.
- Validation Performance: Achieved a peak
val-aux/sciknoweval/reward/mean@16score of 53.50% during validation, indicating strong performance in evaluating biological scientific knowledge.
Intended Use Cases
This model is particularly well-suited for applications requiring a deep understanding of biological concepts and data. Developers can consider this model for:
- Scientific Information Extraction: Extracting key facts and relationships from biological texts.
- Biology-specific Question Answering: Answering queries within the domain of biology.
- Knowledge Evaluation: Assessing and generating content related to biological sciences.
It's important to note that the uploaded weights correspond to the final global_step_100/actor checkpoint, and while the best validation performance was observed at step 90, the final checkpoint is provided.