surrey-nlp/diallm-llama-dpo-brit
The surrey-nlp/diallm-llama-dpo-brit is an 8 billion parameter Llama 3.1-based language model developed by Surrey NLP, specifically adapted for Northern British English. It was continually pretrained on the International Corpus of English and fine-tuned using Direct Preference Optimization (DPO) with dialect-specific preference data. This model excels at generating text in a specific English dialect, making it suitable for applications requiring nuanced regional linguistic output.
Loading preview...
Overview
The surrey-nlp/diallm-llama-dpo-brit is an 8 billion parameter language model built upon the Llama 3.1 architecture, developed by Surrey NLP. This model is a key component of the "DiaLLM: An Investigation into the Robustness-Generation Gap in English Dialect Adaptation" research. Its primary focus is on adapting to and generating text in Northern British English (en-UK).
Key Adaptation Details
- Base Model: Llama 3.1-8B.
- Target Variety: Specifically adapted for en-UK (Northern British English).
- Adaptation Process: Involved continual pretraining on the International Corpus of English (~20M tokens across 18 varieties), followed by dialect-specific Supervised Fine-Tuning (SFT) on Multi-VALUE-transformed en-UK preference data. The final alignment was achieved using Direct Preference Optimization (DPO) with target-variety preference pairs.
- Training Framework: Fine-tuned using TRL, building upon a previous SFT version.
Use Cases
- Dialect-Specific Text Generation: Ideal for applications requiring text output that accurately reflects Northern British English linguistic nuances.
- Linguistic Research: Useful for studying dialect adaptation in large language models and investigating the robustness-generation gap.
- Content Localization: Can be employed for localizing content to specific English dialects.