Encoder comparison after equivalent retrieval fine-tuning
Collection
ModernBERT, Ettin and UltraBERT: 54 retrieval models, frozen data and results. Validation-selected models first. Matched training, one seed. • 59 items • Updated • 4
Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('parquet', {}), NamedSplit('validation'): ('json', {})}
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Encoder comparison: frozen retrieval data
Results, plots, logs and experiment artifacts have moved to the public bucket.
This repository retains the fixed inputs used to compare ModernBERT, Ettin and UltraBERT after dense, multi-vector and reranker fine-tuning.
- train.json.gz: 32,768 MS MARCO hard-negative triplets.
- validation.json.gz: 1,024 disjoint queries with fixed candidates.
- tests.json.gz: NanoNQ, NanoFiQA2018 and NanoSciFact inputs, qrels and BM25 candidates.
- manifest.json: source revisions, sampling details and content hashes.
The experiment uses frozen data revision 0d882e8af96100b875a30b78de98ea8efe99e4db. Original dataset terms apply.
Archived experiment snapshot preserves the original scripts and artifact history. Finished models remain in their model repositories.
- Downloads last month
- 6,173