HYU-NLP-EVAL/qwen3-1.7b-rar-science-static-r0-step-000

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 3, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The HYU-NLP-EVAL/qwen3-1.7b-rar-science-static-r0-step-000 is a 1.7 billion parameter Qwen3-based causal language model, developed by HYU-NLP-EVAL, specifically fine-tuned using the GRPO algorithm for research into reward saturation and static-rubric staleness in the science domain. This model serves as a research artifact, providing a policy checkpoint from a static-rubric discriminability-horizon experiment. It is designed for studying policy optimization dynamics rather than general-purpose application, with a context length of 32768 tokens.

Loading preview...

Overview

This model, HYU-NLP-EVAL/qwen3-1.7b-rar-science-static-r0-step-000, is a 1.7 billion parameter Qwen3-based causal language model. It represents a specific policy checkpoint from a static-rubric discriminability-horizon experiment conducted by HYU-NLP-EVAL. The model was fine-tuned using the GRPO (Generalized Policy Optimization) algorithm, with training focused on a frozen prompt-specific initial rubric (R0) within the science domain.

Key Characteristics

  • Base Model: Qwen/Qwen3-1.7B
  • Fine-tuning Algorithm: GRPO (Generalized Policy Optimization)
  • Training Reward: Frozen prompt-specific initial rubric (R0)
  • Domain: Science, specifically for research into reward saturation and static-rubric staleness.
  • Export Format: Hugging Face Transformers, BF16 safetensors, including model weights, configuration, tokenizer, and chat template.
  • Context Length: 32768 tokens.

Intended Use

This model is primarily a research artifact designed for studying the dynamics of reward saturation and the staleness of static rubrics during policy optimization. It is not intended for general-purpose applications or as a production-ready model. The science training for this specific experiment was saved through step 3, with this repository containing a checkpoint from step 0.