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.
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 Llama2-13B, 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.
