README.md

metadata

license: mit
datasets:
  - miracl/miracl
language:
  - fr
metrics:
  - precision
base_model:
  - antoinelouis/crossencoder-camembert-base-mmarcoFR

MIRACL Cross-Encoder (fr)

This model is a fine-tuned version of antoinelouis/crossencoder-camembert-base-mmarcoFR on the MIRACL dataset for fr language. It uses hard negative mining with BM25 for better training data.

Training

• The model was trained on MIRACL fr dataset
• Hard negative mining was performed using BM25
• For each query, we used all positive passages and up to 30 negative passages (combination of original negatives and BM25 hard negatives)

Usage

from sentence_transformers.cross_encoder import CrossEncoder

model = CrossEncoder("azat-serikbayev/crossencoder-camembert-base-mmarcoFR-miracl-fr")
scores = model.predict([["query", "document_text"]])