HYU-NLP-EVAL/qwen3-1.7b-rar-medicine-static-r0-step-003
The HYU-NLP-EVAL/qwen3-1.7b-rar-medicine-static-r0-step-003 model is a 1.7 billion parameter Qwen3-based causal language model developed by HYU-NLP-EVAL. This specific policy checkpoint is a research artifact from an experiment studying reward saturation and static-rubric staleness during policy optimization, specifically within the medical domain. It is fine-tuned using the GRPO algorithm with a frozen prompt-specific initial rubric (R0) and is intended for research into policy optimization dynamics.
Loading preview...
Overview
This repository hosts a policy checkpoint from the HYU-NLP-EVAL's static-rubric discriminability-horizon experiment, specifically qwen3-1.7b-rar-medicine-static-r0-step-003. It is based on the Qwen/Qwen3-1.7B model and was trained using the GRPO (Generalized Reward Policy Optimization) algorithm. The training utilized a frozen prompt-specific initial rubric (R0) within the medical domain.
Key Characteristics
- Base Model: Qwen3-1.7B, a 1.7 billion parameter causal language model.
- Training Algorithm: GRPO, focused on policy optimization.
- Reward Mechanism: Employs a static, prompt-specific initial rubric (R0).
- Domain: Specialized for the medical domain, covering ten planned audit points through step 48 of the experiment.
- Export Format: Provided in Hugging Face Transformers format with BF16 safetensors, including model weights, configuration, tokenizer, and chat template.
Intended Use
This checkpoint is primarily a research artifact designed for studying reward saturation and the staleness of static rubrics during policy optimization. It is explicitly stated that these medicine checkpoints are not medical devices and should not be used as a substitute for professional medical advice. Developers can load the model using AutoModelForCausalLM and AutoTokenizer from the transformers library.