minchaoh2002/Qwen3-8B-PragReST
minchaoh2002/Qwen3-8B-PragReST is a Qwen3-8B causal language model developed by Minchao Huang and Jihyung Park, fine-tuned using the PragReST framework. This model specializes in pragmatic language understanding and counterfactual pragmatic reasoning, addressing implied meaning, speaker intent, and social context. It achieves strong performance on benchmarks like PragMega, Ludwig, MetoQA, and AltPrag, making it suitable for tasks requiring deep linguistic interpretation.
Loading preview...
Overview
minchaoh2002/Qwen3-8B-PragReST is a Qwen3-8B model developed by Jihyung Park and Minchao Huang from The University of Texas at Austin, specifically trained with PragReST (Pragmatic Reasoning via Self-Training). This framework is designed to enhance a model's ability to understand pragmatic language, focusing on implied meanings, speaker intent, implicature, presupposition, metonymy, and social context that are not explicitly stated in the text.
Key Capabilities
- Pragmatic Language Understanding: Excels at interpreting non-explicit linguistic nuances.
- Counterfactual Pragmatic Reasoning: Trained to reason about scenarios where the intended meaning deviates from the literal.
- Self-Supervised Training: Utilizes a unique self-supervised framework involving supervised fine-tuning with counterfactual bootstrapping, followed by GRPO reinforcement learning.
Performance
The model demonstrates improved performance over the base Qwen3-8B Instruct model on several pragmatic reasoning benchmarks:
- PragMega: Achieves 79.29% (vs. 73.37% for base).
- Ludwig: Achieves 83.33% (vs. 80.33% for base).
- MetoQA: Achieves 80.72% (vs. 73.52% for base).
- AltPrag: Achieves 7.62% (vs. 7.24% for base).
Good For
- Applications requiring deep understanding of human communication, including implied meanings and social context.
- Research and development in pragmatic AI and natural language understanding.
- Tasks involving complex linguistic interpretation beyond literal text.