README.md

---
license: apache-2.0
dataset_info:
- config_name: arabic_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 338926883
  dataset_size: 535572160
- config_name: chinese_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 339423173
  dataset_size: 535572160
- config_name: dutch_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 338910473
  dataset_size: 535572160
- config_name: english_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 339859952
  dataset_size: 535572160
- config_name: french_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 339387465
  dataset_size: 535572160
- config_name: german_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 338132277
  dataset_size: 535572160
- config_name: hindi_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 339380999
  dataset_size: 535572160
- config_name: indonesian_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 339703081
  dataset_size: 535572160
- config_name: italian_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 339400584
  dataset_size: 535572160
- config_name: japanese_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 337734717
  dataset_size: 535572160
- config_name: portuguese_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 339626925
  dataset_size: 535572160
- config_name: russian_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 338596435
  dataset_size: 535572160
- config_name: sentence-transformers-msmarco-hard-negatives
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  splits:
  - name: train
    num_bytes: 846201448
    num_examples: 502939
  download_size: 662470387
  dataset_size: 846201448
- config_name: sentence-transformers-msmarco-hard-negatives-bm25
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  splits:
  - name: train
    num_bytes: 213472888
    num_examples: 502912
  download_size: 176954469
  dataset_size: 213472888
- config_name: spanish_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 535572160
    num_examples: 502912
  download_size: 339556153
  dataset_size: 535572160
- config_name: vietnamese_bge-reranker-v2-m3
  features:
  - name: qid
    dtype: int64
  - name: pos
    sequence: int64
  - name: neg
    sequence: int64
  - name: pos.score
    sequence: float64
  - name: neg.score
    sequence: float64
  splits:
  - name: train
    num_bytes: 278081216
    num_examples: 502912
  download_size: 185789847
  dataset_size: 278081216
configs:
- config_name: arabic_bge-reranker-v2-m3
  data_files:
  - split: train
    path: arabic_bge-reranker-v2-m3/train-*
- config_name: chinese_bge-reranker-v2-m3
  data_files:
  - split: train
    path: chinese_bge-reranker-v2-m3/train-*
- config_name: dutch_bge-reranker-v2-m3
  data_files:
  - split: train
    path: dutch_bge-reranker-v2-m3/train-*
- config_name: english_bge-reranker-v2-m3
  data_files:
  - split: train
    path: english_bge-reranker-v2-m3/train-*
- config_name: french_bge-reranker-v2-m3
  data_files:
  - split: train
    path: french_bge-reranker-v2-m3/train-*
- config_name: german_bge-reranker-v2-m3
  data_files:
  - split: train
    path: german_bge-reranker-v2-m3/train-*
- config_name: hindi_bge-reranker-v2-m3
  data_files:
  - split: train
    path: hindi_bge-reranker-v2-m3/train-*
- config_name: indonesian_bge-reranker-v2-m3
  data_files:
  - split: train
    path: indonesian_bge-reranker-v2-m3/train-*
- config_name: italian_bge-reranker-v2-m3
  data_files:
  - split: train
    path: italian_bge-reranker-v2-m3/train-*
- config_name: japanese_bge-reranker-v2-m3
  data_files:
  - split: train
    path: japanese_bge-reranker-v2-m3/train-*
- config_name: portuguese_bge-reranker-v2-m3
  data_files:
  - split: train
    path: portuguese_bge-reranker-v2-m3/train-*
- config_name: russian_bge-reranker-v2-m3
  data_files:
  - split: train
    path: russian_bge-reranker-v2-m3/train-*
- config_name: sentence-transformers-msmarco-hard-negatives
  data_files:
  - split: train
    path: sentence-transformers-msmarco-hard-negatives/train-*
- config_name: sentence-transformers-msmarco-hard-negatives-bm25
  data_files:
  - split: train
    path: sentence-transformers-msmarco-hard-negatives-bm25/train-*
- config_name: spanish_bge-reranker-v2-m3
  data_files:
  - split: train
    path: spanish_bge-reranker-v2-m3/train-*
- config_name: vietnamese_bge-reranker-v2-m3
  data_files:
  - split: train
    path: vietnamese_bge-reranker-v2-m3/train-*
---

- `hotchpotch/mmarco-hard-negatives-reranker-score`

This repository contains data from [mMARCO](https://huggingface.co/datasets/unicamp-dl/mmarco) scored using the reranker [BAAI/bge-reranker-v2-m3](https://huggingface.co/BAAI/bge-reranker-v2-m3).

## Languages Covered

```
target_languages = ["english", "chinese", "french", "german", "indonesian", "italian", "portuguese", "russian", "spanish", "arabic", "dutch", "hindi", "japanese", "vietnamese"]
```

## Hard Negative Data

The hard negative data is derived from [sentence-transformers-msmarco-hard-negatives-bm25](https://huggingface.co/datasets/sentence-transformers/msmarco-hard-negatives), randomly sampling 32 or 64 instances for use.

## License

This project adheres to the same license as mMARCO: **Apache License 2.0**.

# Example


```python
# target languages => ["english", "chinese", "french", "german", "indonesian", "italian", "portuguese", "russian", "spanish", "arabic", "dutch", "hindi", "japanese", "vietnamese"]

lang = "spanish"
repo_id = "hotchpotch/mmarco-hard-negatives-reranker-score"
reranker = "bge-reranker-v2-m3"
subset = f"{lang}_{reranker}"
mapping = f"mappings/{lang}_joblib.pkl.gz"

from datasets import load_dataset
import joblib
from huggingface_hub import hf_hub_download

queries_ds = load_dataset(
    "unicamp-dl/mmarco", "queries-" + lang, split="train", trust_remote_code=True
)
collection_ds = load_dataset(
    "unicamp-dl/mmarco",
    "collection-" + lang,
    split="collection",
    trust_remote_code=True,
)
score_ds = load_dataset(repo_id, subset, split="train")


mapping_file = hf_hub_download(repo_type="dataset", repo_id=repo_id, filename=mapping)

index_mapping_dict = joblib.load(mapping_file)
query_id_dict = index_mapping_dict["query_id_dict"]
collection_id_dict = index_mapping_dict["collection_id_dict"]


def get_query_text(query_id) -> str:
    idx = query_id_dict[query_id]
    return queries_ds[idx]["text"]  # type: ignore


def get_collection_text(doc_id) -> str:
    idx = collection_id_dict[doc_id]
    return collection_ds[idx]["text"]  # type: ignore

for i in range(5):
    qid: int = score_ds[i]["qid"]
    pos: list[int] = score_ds[i]["pos"]
    pos_score: list[float] = score_ds[i]["pos.score"]
    neg: list[int] = score_ds[i]["neg"]
    neg_score: list[float] = score_ds[i]["neg.score"]
    query = get_query_text(qid)
    pos_docs = [get_collection_text(doc_id)[0:64] for doc_id in pos]
    neg_docs = [get_collection_text(doc_id)[0:64] for doc_id in neg]
    print(f"# Query: {query}")
    print("## Positive docs:")
    for doc, score in zip(pos_docs, pos_score):
        print(f"  {doc} ({score})")
    print("## Negative docs:")
    for doc, score in list(zip(neg_docs, neg_score))[0:5]:
        print(f"  {doc} ({score})")
    print("-------")
```

output

```
# Query: ¿Qué son las artes liberales?
## Positive docs:
  Artes liberales. 1. el curso académico de instrucción en una uni (0.99770385)
## Negative docs:
  Grandes Ligas. Puede elegir entre una variedad de especializacio (0.69760895)
  BA = Licenciatura en Artes BS = Licenciatura en Ciencias Creo qu (0.24364243)
  ¿Qué es una Licenciatura en Artes (B.A.)? Un programa de licenci (0.20641373)
  ¿Qué significa LCSW? / Human and Social ... / Liberal Arts y ... (0.0140636265)
  definición de artes liberales Las áreas de aprendizaje que culti (0.9963924)
-------
# Query: ¿Cuál es el mecanismo de acción de los fármacos fibrinolíticos o trombolíticos?
## Positive docs:
  Hematología clínica de BailliÃÆ'¨re. 6 Mecanismo de acción d (0.966347)
## Negative docs:
  Definición y ejemplos de mecanismos de acción. Más en Trastorno  (0.3598139)
  ¿Qué es losartán y cómo funciona (mecanismo de acción)? ¿Qué mar (0.0031480708)
  ActivaseÃ⠀ šÃ,® Una propietaria trombolítico, que puede à ¢  (0.83237296)
  La terapia fibrinolítica, también llamada a veces "terapia tromb (0.92590266)
  Diazepam Valium Mecanismo de acción Valium Mecanismo de acción E (0.040162582)
-------
# Query: ¿Qué es el recuento normal de plataformas?
## Positive docs:
  78 seguidores. R. Las plaquetas son glóbulos diminutos que ayuda (0.10105592)
## Negative docs:
  ¿Qué es la trombocitopenia (recuento bajo de plaquetas)? Las pla (0.047337715)
  Calificación Más reciente Más antiguo. Mejor respuesta: Nancy: e (0.03179867)
  1 Tarifas de solicitud ࢠ€Â⠀ œ $ 80 por una plataform (0.001044386)
  Conteo sanguíneo de MCV. Mi recuento sanguíneo de MCV está en 98 (0.011115014)
  ¿Cuáles son los niveles normales de hemograma para una mujer adu (0.015247591)
-------
# Query: promedio de costo en dólares explicado
## Positive docs:
  El promedio del costo en dólares es simplemente un método para c (0.96771675)
## Negative docs:
  El promedio del costo en dólares es una técnica simple que le pe (0.9859364)
  Anteriormente en Free from Broke, Glen ha abordado el tema del c (0.89274967)
  (TMFMathGuy). 19 de noviembre de 2014 a las 2:15 p.m. El promedi (0.98421544)
  Comprar acciones por valor de $ 2,000 en el primer mes, a $ 14.2 (0.36624968)
  DEFINICIÓN de 'Valor Promedio'. Una estrategia de inversión que  (0.3941427)
-------
# Query: alimentos que ayudan a combatir la gota
## Positive docs:
  Además de seguir una dieta bien balanceada para promover la salu (0.9874721)
## Negative docs:
  Alimentos para la gota: anacardos y gota. Los anacardos y la got (0.84606963)
  Alimentos que debe evitar si tiene gota La gota es un tipo de ar (0.016979992)
  El puerto causa gota. Estos son solo algunos tratamientos natura (0.30850264)
  20 alimentos ricos en purina que debe evitar: alterar su dieta p (0.4324828)
  Alimentos que combaten el cáncer por BÃÆ' © liveau & Gingras (0.010408314)
-------

```