AIJian/PaTaRM-14B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:14BQuant:FP8Ctx Length:32kPublished:Mar 20, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

PaTaRM-14B is a 14 billion parameter language model developed by Ai Jian, based on the Qwen3 architecture with a 32768 token context length. This model is part of the PaTaRM series, which focuses on Preference-Aware Task-Adaptive Reward Modeling. It is designed to bridge pairwise and pointwise signals, making it suitable for tasks requiring nuanced preference understanding and reward-based optimization.

Loading preview...

PaTaRM-14B: Preference-Aware Task-Adaptive Reward Modeling

PaTaRM-14B is a 14 billion parameter model from the PaTaRM series, developed by Ai Jian. It is built upon the Qwen3-14B base architecture and features a substantial context length of 32768 tokens. The core innovation behind PaTaRM models lies in their approach to "Preference-Aware Task-Adaptive Reward Modeling," which aims to effectively integrate both pairwise and pointwise signals during training.

Key Capabilities

  • Preference-Aware Learning: Designed to understand and leverage user preferences more effectively by bridging different types of feedback signals.
  • Task-Adaptive Reward Modeling: Incorporates a mechanism for reward modeling that adapts to specific tasks, potentially leading to more aligned and performant outputs.
  • Qwen3 Base: Benefits from the robust capabilities and architectural strengths of the Qwen3 foundation model.

Good for

  • Applications requiring sophisticated preference learning and reward-based optimization.
  • Research and development in advanced language model alignment techniques.
  • Tasks where understanding nuanced user feedback is critical for performance improvement.