RankVicuna (No Data Augmentation) Overview
RankVicuna is a 7 billion parameter chat assistant developed by Castorini, fine-tuned from the Llama 2 base model. This specific version, v1_noda, is distinguished by being trained without additional data augmentation, focusing on a direct fine-tuning approach from lmsys/vicuna-7b-v1.5 using supervised instruction fine-tuning. The model's architecture is an auto-regressive language model based on the transformer design.
Key Capabilities & Focus
- Research in LLMs and Retrieval: Primarily designed for academic and experimental work exploring the synergy between large language models and information retrieval systems.
- Llama 2 Foundation: Benefits from the robust architecture and pre-training of the Llama 2 model family.
- Instruction Fine-tuning: Utilizes supervised instruction fine-tuning on user-shared conversations from ShareGPT to enhance its chat assistant capabilities.
Intended Use Cases
- Information Retrieval Research: Ideal for researchers investigating how LLMs can improve ranking, relevance, and search functionalities.
- Natural Language Processing Experiments: Suitable for hobbyists and researchers working on various NLP tasks, particularly those involving conversational data and retrieval contexts.
- Evaluation on Standard Benchmarks: The model has been evaluated on datasets like DL19/DL20, with further details available in its associated paper.