notoookay/ragler-llama2-7b-reranker

Update (Dec 31, 2024):

We have updated our model.

1. The previous version uses ms-marco-MiniLM-L-6-v2 as base model, the updated version takes ms-marco-MiniLM-L-12-v2 as base model which shows better generality and performance improvements.

2. The new model uses confidence threshold of 0.8 compared to prior model without confidence-based training.

This model is the corresponding re-ranker supervised by instruction-tuned Llama2-7B, check our paper for more details.

The model generates a relevance score directly during ranking. Please replace the re-ranker argument in the scripts with model id.