stefan-jo/scifact-train-mined-reranker-scores

SciFact Train Mined Reranker Scores

This repository contains a derived SciFact training dataset for retrieval and distillation experiments. It was constructed from the SciFact training split by mining candidate negatives and then assigning teacher scores with a reranker.

Dataset Summary

• Source dataset: SciFact train split in BEIR-style preprocessing
• Released split: train
• Number of rows: 809
• Row schema:
  • query_id
  • document_ids
  • scores
  • labels

Each final row contains a ranked candidate set with:

• 1 positive document
• 15 mined negative documents
• 16 scored documents total

Data Construction Process

The mining and scoring pipeline follows the scripts in examples/pooling/train/data/ and the paper Learn to Pool: Lightweight Fine-Tuning for Flexible Multi-Vector Compression.

For each query, positive documents are taken from the SciFact train qrels. A pool of 100 hard negative candidates is then mined with BAAI/bge-small-en-v1.5, and an additional pool of 30 random negatives is sampled from the remaining corpus.

All candidate query-document pairs are then scored with BAAI/bge-reranker-v2-gemma. Following the filtering strategy described in the paper, negatives with a reranker score higher than 95% of the positive anchor score for that query are removed to reduce the chance of keeping false negatives.

The final per-query candidate set is built from the remaining documents by keeping the top positive, the top-scoring hard negatives, additional medium negatives sampled from the remaining hard-negative pool, and random negatives sampled from the random pool. This yields a compact scored tuple with exactly 16 documents per query: 1 positive and 15 mined negatives spanning hard, medium, and random examples.

Released Features

The released train split stores:

• query_id: query identifier
• document_ids: ranked list of document identifiers
• scores: reranker scores aligned with document_ids
• labels: document roles aligned with document_ids and scores, one of:
  • positive
  • hard_negative
  • medium_negative
  • random_negative

The labels column indicates how each document entered the final candidate set. This is useful if you want to distinguish stronger mined negatives from easier random negatives during analysis or downstream training.

This release contains ID-based training annotations, not the full query and document texts. The query_id and document_ids fields are meant to be resolved against the corresponding SciFact train split in BEIR format.

In the training pipeline, this is done by loading the BEIR train split, building:

• a query_id -> text mapping from the query set
• a document_id -> text mapping from the corpus

and then applying those mappings to the scored tuples during training. See examples/pooling/train/train_beir_colbert_distillation.py for a concrete example using load_beir_train_split(...) together with utils.KDProcessing(...).

Intended Use

This dataset is intended for:

• retrieval training
• knowledge distillation / teacher-student training
• studying mined and reranker-scored candidate sets
• reproducing the paper's SciFact experiments

Provenance

The released dataset was built using:

• source data: SciFact train split
• mining model: BAAI/bge-small-en-v1.5
• reranker: BAAI/bge-reranker-v2-gemma

License and Data Sources

This repository is released under a mixed-source custom notice because it is derived from SciFact data that combines multiple upstream sources.

Relevant upstream components:

• SciFact claims and evidence annotations: CC BY 4.0
• SciFact corpus source: S2ORC / ODC-By 1.0
• mining model: BAAI/bge-small-en-v1.5 (MIT)
• reranker: BAAI/bge-reranker-v2-gemma (Apache-2.0)

Users should review and comply with the upstream attribution and source terms in addition to this repository notice.

References

• Paper: Learn to Pool: Lightweight Fine-Tuning for Flexible Multi-Vector Compression
• Code: stefan-jo/pylate