mjaliz/bslm-pair-206k-data

BSLM Pair 206k Query-Title Dataset

Query/product-title pair dataset used to fine-tune mjaliz/bslm-pair-206k with Sentence Transformers MultipleNegativesRankingLoss.

The dataset was built from ecommerce query-to-anchor candidates by selecting one positive product title per query (top1) to reduce false negatives in in-batch negative training.

Files

File	Rows	Purpose
train.jsonl	202,149	Pair training rows
dev.jsonl	2,063	Pair validation rows
test.jsonl	2,063	Pair test rows
dev_queries.jsonl	2,063	BEIR-style dev queries
dev_corpus.jsonl	2,047	BEIR-style dev corpus
dev_qrels.jsonl	2,063	BEIR-style dev relevance labels
test_queries.jsonl	2,063	BEIR-style test queries
test_corpus.jsonl	2,046	BEIR-style test corpus
test_qrels.jsonl	2,063	BEIR-style test relevance labels
stats.json	1	Dataset build statistics

Main pair schema

The main train.jsonl, dev.jsonl, and test.jsonl files contain one JSON object per line with fields such as:

• query_id: deterministic query identifier
• anchor: query text
• positive: positive product title
• product_id: positive product identifier
• anchor_score: candidate score
• anchor_rank: selected candidate rank
• anchor_margin: margin from the next candidate where available
• query_frequency: source query frequency
• mapping_confidence: source mapping confidence
• mapping_method: source mapping method

For Sentence Transformers training, only anchor and positive are required.

Loading example

from datasets import load_dataset

pairs = load_dataset("mjaliz/bslm-pair-206k-data")
print(pairs["train"][0]["anchor"])
print(pairs["train"][0]["positive"])

Training context

This dataset corresponds to the pair-training stage described in the repository README:

CUDA_VISIBLE_DEVICES=2,3,5 torchrun --standalone --nproc_per_node=3 scripts/train_biencoder.py \
  --model-name-or-path mjaliz/bslm-mlm-25M-ptdrw \
  --train-file data/pairs/train.jsonl \
  --dev-file data/pairs/dev.jsonl \
  --dev-queries data/pairs/dev_queries.jsonl \
  --dev-corpus data/pairs/dev_corpus.jsonl \
  --dev-qrels data/pairs/dev_qrels.jsonl \
  --output-dir models/bslm-mnrl-query-title \
  --loss mnrl

Notes

• Splits are by query_id to avoid leakage across train/dev/test.
• The pair builder selected one positive per query (positive_selection=top1).
• Source confidences included high and medium.
• License is not specified here; verify downstream usage permissions before redistributing derived artifacts.