minchaoh2002/Qwen3-14B-PragReST
minchaoh2002/Qwen3-14B-PragReST is a 14 billion parameter causal language model based on the Qwen3 architecture, developed by Jihyung Park, Minchao Huang, Leqi Liu, and Elias Stengel-Eskin. This model is specifically fine-tuned using the PragReST (Pragmatic Reasoning via Self-Training) framework to enhance pragmatic language understanding. It excels at interpreting implied meaning, speaker intent, implicature, and other complex linguistic nuances not explicitly stated in text. The model is designed for advanced pragmatic reasoning tasks, making it suitable for applications requiring deep contextual and social understanding.
Loading preview...
Overview
minchaoh2002/Qwen3-14B-PragReST is a 14 billion parameter model built upon the Qwen3-14B base, developed by researchers at The University of Texas at Austin. Its core innovation lies in its training with PragReST (Pragmatic Reasoning via Self-Training), a self-supervised framework designed to significantly improve pragmatic language understanding.
Key Capabilities
- Pragmatic Language Understanding: The model is specifically trained to interpret implied meanings, speaker intent, implicature, presupposition, metonymy, and social context, going beyond literal text interpretation.
- Counterfactual Pragmatic Reasoning: Utilizes a training methodology involving supervised fine-tuning with counterfactual bootstrapping and GRPO reinforcement learning to enhance its reasoning abilities.
- Benchmark Performance: Demonstrates improved performance over the base Qwen3-14B Instruct model across various pragmatic reasoning benchmarks, including PragMega, Ludwig, MetoQA, and AltPrag. For instance, it achieves 85.80 on PragMega and 86.50 on Ludwig, outperforming the base model.
Good For
- Advanced NLP Applications: Ideal for tasks requiring a deep understanding of human communication nuances, where implied meaning is crucial.
- Research in Pragmatics: A valuable tool for researchers exploring pragmatic reasoning, implicature, and social context in language models.
- Contextual AI Systems: Suitable for systems that need to infer non-explicit information from conversations or text, such as sophisticated chatbots or dialogue agents.
This model is associated with the paper "PragReST: Self-Reinforcing Counterfactual Reasoning for Pragmatic Language Understanding" (arXiv:2606.18624).