bobox/natural-questions-bge-m3-test2

Processed Dataset with Multi-Vector BGE-M3 Scores

This dataset was automatically built using a custom pipeline to format data for Sentence Transformers training.

Source Dataset

• Original Dataset: sentence-transformers/natural-questions
• Split processed: train
• Anchor Column: query
• Positive Column: answer
• Negative Column: negative

Hard Negative Mining

Hard negatives were aggressively mined from the dataset to improve model training robustness.

• Bi-Encoder (Mining Model): Snowflake/snowflake-arctic-embed-l-v2.0

• Cross-Encoder (Reranker): None Used

• Negatives per positive: 1

• Relative Margin applied: 0.06

• Absolute Margin applied: 0.04

• range min/max: 0 - 100

• negative score min/max: 0.1 - 0.9

• Sampling strategy: top

• FAISS used: True

Detailed Multi-Vector Scoring (BAAI/bge-m3)

The dataset includes granular similarity scores computed using BAAI/bge-m3. These scores include Dense, Sparse (Lexical), and ColBERT-style (Multi-vector) representations, plus an aggregate score natively weighted as 0.4 * dense + 0.2 * sparse + 0.4 * colbert.

• Scoring Model: BAAI/bge-m3
• Max Passage Length: 4000

New Features Added:

• anchor_pos_aggregate_sim_m3, anchor_pos_list_sim_m3,

• anchor_neg_aggregate_sim_m3, anchor_neg_list_sim_m3,

• scores: Bi-Encoder rescoring predictions from the hard-negative mining phase.