HYU-NLP-EVAL/qwen3-1.7b-rar-science-static-r0-step-003
HYU-NLP-EVAL/qwen3-1.7b-rar-science-static-r0-step-003 is a 2 billion parameter Qwen3-1.7B model, fine-tuned using the GRPO algorithm with a frozen prompt-specific initial rubric (R0) for the Science domain. This model serves as a research artifact for studying reward saturation and static-rubric staleness during policy optimization. It is specifically designed for research into reinforcement learning policies and their evolution over training steps.
Loading preview...
Model Overview
This repository hosts a policy checkpoint from the static-rubric discriminability-horizon experiment, specifically qwen3-1.7b-rar-science-static-r0-step-003. It is based on the Qwen/Qwen3-1.7B model and has been fine-tuned using the GRPO (Generalized Reinforcement Policy Optimization) algorithm.
Key Characteristics
- Base Model: Qwen/Qwen3-1.7B (2 billion parameters).
- Fine-tuning: Utilizes the GRPO algorithm with a frozen prompt-specific initial rubric (
R0). - Domain: Specialized for the Science domain, as indicated in its name.
- Purpose: Primarily intended as a research artifact to investigate reward saturation and the staleness of static rubrics during policy optimization.
- Checkpoint Details: Contains model weights, configuration, tokenizer, and chat template in BF16 safetensors format. Excludes optimizer, scheduler, trainer state, rollouts, rubrics, and evaluation data.
Intended Use
This model is a research artifact for studying the dynamics of reinforcement learning policies. It is not designed for general-purpose applications but rather for specific research into how reward functions and training steps influence model performance and stability in a scientific context. The model represents an early training step (step 3) within the experiment.