tevatron-elastic-bert-reranker-depth

A reranker trained with Tevatron-Elastic, which trains one checkpoint to serve many operating points along the depth / width / token compression axes. This checkpoint is an elastic depth axis (early exit): one checkpoint serves several layer counts.

• Base model: google-bert/bert-base-uncased
• Task: reranker
• Elastic axis: depth
• Training data: rlhn/rlhn-680K, max length 512, query:/passage: prefixes.

Full-point BEIR-15 nDCG@10: 0.453.

Load with the Tevatron-Elastic framework and select an operating point with prune_to / encode_at; see the repository for usage. Part of a release of 20 checkpoints (3 backbones, retrieval and reranking, all compression axes) accompanying the Tevatron-Elastic paper. Reported as a reproducibility resource, not a state-of-the-art claim.