TMLR-Group-HF/Co-rewarding-I-Qwen3-8B-Base-DAPO14k

Model Overview

This model, TMLR-Group-HF/Co-rewarding-I-Qwen3-8B-Base-DAPO14k, is an 8 billion parameter Qwen3-Base architecture. It was developed by Co-rewarding-I and fine-tuned on the DAPO-14k dataset. The core innovation lies in its training methodology, which utilizes the Co-rewarding-I framework.

Key Capabilities

• Enhanced Reasoning: The model is specifically designed to improve reasoning abilities in large language models (LLMs).
• Stable Self-supervised RL: It employs Co-rewarding-I, a data-side instantiation of a novel self-supervised reinforcement learning (RL) framework. This approach aims to provide more stable training compared to traditional single-view self-rewarding methods.
• Mitigation of Training Issues: Co-rewarding-I addresses common problems like training collapse and reward hacking by deriving reward signals from contrastive agreement across semantically analogous questions.
• Mathematical Reasoning: The framework is particularly effective for complex challenges such as mathematical reasoning.

What Makes This Model Different?

This model stands out due to its Co-rewarding-I training framework, which is detailed in the paper "Co-rewarding: Stable Self-supervised RL for Eliciting Reasoning in Large Language Models" [https://huggingface.co/papers/2508.00410]. Unlike many other LLMs, its training focuses on a robust self-supervised RL approach to elicit and stabilize reasoning capabilities, rather than relying solely on supervised fine-tuning or simpler RL methods. This makes it particularly suitable for applications requiring reliable and consistent reasoning performance.