surrey-nlp/diallm-qwen-grpo-aus
The surrey-nlp/diallm-qwen-grpo-aus model is an 8 billion parameter Qwen 3-8B variant, fine-tuned for Australian English (en-AU) dialect adaptation. It was continually pretrained on the International Corpus of English and further adapted using dialect-specific Supervised Fine-Tuning (SFT) on Multi-VALUE-transformed en-AU preference data. The model employs Generative Reinforcement Learning from Policy Optimization (GRPO) with target-variety preference pairs to enhance its robustness and generation capabilities for Australian English.
Loading preview...
DiaLLM - Qwen 3-8B - Australian English Adaptation
This model, surrey-nlp/diallm-qwen-grpo-aus, is an 8 billion parameter variant of the Qwen 3-8B base model, specifically adapted for Australian English (en-AU). It is a key component of the "DiaLLM: An Investigation into the Robustness-Generation Gap in English Dialect Adaptation" research.
Key Adaptation Details
- Target Variety: Australian English (en-AU).
- Adaptation Method: The model underwent an "explicit" (variety-targeted) adaptation thread.
- Training Process:
- Initially continually pretrained on the International Corpus of English (18 varieties, ~20M tokens).
- Further adapted via dialect-specific Supervised Fine-Tuning (SFT) using Multi-VALUE-transformed en-AU preference data.
- Aligned using Generative Reinforcement Learning from Policy Optimization (GRPO) with target-variety preference pairs.
- Base Model: It is a fine-tuned version of
jordanpainter/diallm-qwen-sft-aus.
Use Cases
This model is particularly suited for applications requiring language generation or understanding with a strong emphasis on Australian English linguistic nuances and expressions. It aims to bridge the robustness-generation gap in dialect adaptation, making it valuable for research and development in localized NLP tasks.
Further Resources
- Code, Checkpoints, Datasets: surrey-nlp/diallm GitHub repository
- Research Paper: arXiv:2607.07669