Uploaded January 2024 | Updated September 2026, 27 minutes ago
RankZephyr is an open-source zero-shot listwise reranker model. Pradeep et al. trained the Zepyhr 7B model from Hugging Face which was built on top of Mistral. They found that RankZephyr performs better than RankGPT4 in some instances (evaluated on multiple datasets).
In this video, I cover:
1. Listwise reranking
2. How they get the language model to output a list
3. Sliding window approach to avoid going over the token window
In conclusion, RankZephyr shows the promise of fine-tuning smaller language models to tackle specific tasks with data labeled by larger models!
Paper: arxiv.org/abs/2312.02724
GitHub repo: github.com/castorini/rank_llm
RankZephyr is an open-source zero-shot listwise reranker model. Pradeep et al. trained the Zepyhr 7B model from Hugging Face which was built on top of Mistral. They found that RankZephyr performs better than RankGPT4 in some instances (evaluated on multiple datasets).
In this video, I cover:
1. Listwise reranking
2. How they get the language model to output a list
3. Sliding window approach to avoid going over the token window
In conclusion, RankZephyr shows the promise of fine-tuning smaller language models to tackle specific tasks with data labeled by larger models!
Paper: arxiv.org/abs/2312.02724
GitHub repo: github.com/castorini/rank_llm
