hotchpotch/bge-m3-data-finetune-unified

Mirror of the original Shitao/bge-m3-data corpus, repackaged from gzip-compressed JSONL shards into Hugging Face Dataset confi gs stored as Parquet. Content is unchanged; this repo only standardizes the storage format and centralizes all length buckets un der one dataset namespace.

By using this dataset, which includes up to seven hard negative texts per query (as in the original bge-m3 training data), you can easily fine-tune retrieval models or other search systems.

Provenance and conversion

• Source: Shitao/bge-m3-data, which hosts the official fine-tuning data fo r the BGE-M3 model.
• Formatting changes: Only the container format changed (JSONL → HF Dataset → Parquet on the Hub). Field names and values are id entical to the source.

Schema

Each config keeps the original triplet fields:

• query: string
• pos: list of positive passages (strings)
• neg: list of negative passages (strings)

Using the dataset

from datasets import load_dataset

ds = load_dataset("hotchpotch/bge-m3-data-finetune-unified", "miracl_en_len-0-500")
row = ds["train"][0]

preview = lambda t: t.replace("\n", " ")[:50] + ("..." if len(t) > 50 else "")
print("query:", row["query"])
print("pos[0]:", preview(row["pos"][0]))
for i, n in enumerate(row["neg"][:3]):  # show first 3 negatives
    print(f"neg[{i}]:", preview(n))

Example output:

query: When was quantum field theory developed?
pos[0]: History of quantum field theory The third thread i...
neg[0]: AdS/CFT correspondence In quantum field theory, on...
neg[1]: Condensed matter physics The Sommerfeld model and ...
neg[2]: Quantum configuration space In quantum field theor...

Original Dataset Summary

This depository contains all the fine-tuning data for the bge-m3 model, including:

Dataset	Language
MS MARCO	English
NQ	English
HotpotQA	English
TriviaQA	English
SQuAD	English
COLIEE	English
PubMedQA	English
NLI from SimCSE	English
DuReader	Chinese
mMARCO-zh	Chinese
T2Ranking	Chinese
Law-GPT	Chinese
cMedQAv2	Chinese
NLI-zh	Chinese
LeCaRDv2	Chinese
Mr.TyDi	11 languages
MIRACL	16 languages
MLDR	13 languages

Note: The MLDR dataset here is the handled train set of the MLDR dataset.

For more details, please refer to our paper.

Dataset Structure

Each dataset has been split into multiple files according to the tokenized length of the text (tokenizer of bge-m3, i.e. tokenizer of xlm-roberta-large). For example, the MS MARCO dataset has been split into 8 files: msmarco_len-0-500.jsonl, msmarco_len-500-1000.jsonl, ..., msmarco_len-6000-7000.jsonl, msmarco_len-7000-inf.jsonl. All the files are in the jsonl format. Each line of the file is a json object. The following is an example of the json object:

{"query": str, "pos": List[str], "neg":List[str]}

Citation Information

@misc{bge-m3,
      title={BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation},
      author={Jianlv Chen and Shitao Xiao and Peitian Zhang and Kun Luo and Defu Lian and Zheng Liu},
      year={2024},
      eprint={2402.03216},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}