surrey-nlp/diallm-llama-gspo-aus

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Apr 17, 2026License:llama3.1Architecture:Transformer Featherless Exclusive Cold

The surrey-nlp/diallm-llama-gspo-aus model is an 8 billion parameter Llama 3.1-based language model developed by surrey-nlp, specifically adapted for Australian English (en-AU). It was continually pretrained on the International Corpus of English and fine-tuned using dialect-specific SFT on Multi-VALUE-transformed en-AU preference data, followed by Group Sequence Policy Optimization (GSPO). This model is designed to address the robustness-generation gap in English dialect adaptation, making it suitable for tasks requiring nuanced understanding and generation in Australian English.

Loading preview...

Overview

This model, surrey-nlp/diallm-llama-gspo-aus, is an 8 billion parameter variant of Llama 3.1, specifically engineered 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, presented at EMNLP 2026.

Key Capabilities

  • Dialect Adaptation: Explicitly adapted for Australian English, focusing on improving both robustness and generation quality for this specific dialect.
  • Training Methodology: Utilizes a unique adaptation thread involving dialect-specific Supervised Fine-Tuning (SFT) on en-AU preference data, followed by Group Sequence Policy Optimization (GSPO).
  • Base Model: Built upon the robust Llama 3.1-8B architecture, providing a strong foundation for dialectal nuances.
  • Continual Pretraining: Underwent continual pretraining on the International Corpus of English, encompassing 18 varieties and approximately 20 million tokens.

Good For

  • Applications requiring high-quality text generation and understanding in Australian English.
  • Research into dialectal variations and their impact on large language models.
  • Developers interested in models fine-tuned with GSPO for specific linguistic targets.

For more technical details, including code, checkpoints, and preference datasets, refer to the official GitHub repository and the research paper.