metadata

pipeline_tag: text-classification
language: fr
license: mit
datasets:
  - unicamp-dl/mmarco
metrics:
  - recall
tags:
  - passage-reranking
library_name: sentence-transformers
base_model: camembert-base

crossencoder-distilcamembert-mmarcoFR

This is a cross-encoder model for French. It performs cross-attention between a question-passage pair and outputs a relevance score. The model should be used as a reranker for semantic search: given a query and a set of potentially relevant passages retrieved by an efficient first-stage retrieval system (e.g., BM25 or a fine-tuned dense single-vector bi-encoder), encode each query-passage pair and sort the passages in a decreasing order of relevance according to the model's predicted scores.

Usage

Here are some examples for using the model with Sentence-Transformers, FlagEmbedding, or Huggingface Transformers.

Using Sentence-Transformers

Start by installing the library: pip install -U sentence-transformers. Then, you can use the model like this:

from sentence_transformers import CrossEncoder

pairs = [('Question', 'Paragraphe 1'), ('Question', 'Paragraphe 2') , ('Question', 'Paragraphe 3')]

model = CrossEncoder('antoinelouis/crossencoder-distilcamembert-mmarcoFR')
scores = model.predict(pairs)
print(scores)

Using FlagEmbedding

Start by installing the library: pip install -U FlagEmbedding. Then, you can use the model like this:

from FlagEmbedding import FlagReranker

pairs = [('Question', 'Paragraphe 1'), ('Question', 'Paragraphe 2') , ('Question', 'Paragraphe 3')]

reranker = FlagReranker('antoinelouis/crossencoder-distilcamembert-mmarcoFR')
scores = reranker.compute_score(pairs)
print(scores)

Using HuggingFace Transformers

Start by installing the library: pip install -U transformers. Then, you can use the model like this:

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

pairs = [('Question', 'Paragraphe 1'), ('Question', 'Paragraphe 2') , ('Question', 'Paragraphe 3')]

tokenizer = AutoTokenizer.from_pretrained('antoinelouis/crossencoder-distilcamembert-mmarcoFR')
model = AutoModelForSequenceClassification.from_pretrained('antoinelouis/crossencoder-distilcamembert-mmarcoFR')
model.eval()

with torch.no_grad():
    inputs = tokenizer(pairs, padding=True, truncation=True, return_tensors='pt', max_length=512)
    scores = model(**inputs, return_dict=True).logits.view(-1, ).float()
print(scores)

Evaluation

We evaluate the model on 500 random training queries from mMARCO-fr (which were excluded from training) by reranking subsets of candidate passages comprising of at least one relevant and up to 200 BM25 negative passages for each query. Below, we compare the model performance with other cross-encoder models fine-tuned on the same dataset. We report the R-precision (RP), mean reciprocal rank (MRR), and recall at various cut-offs (R@k).

	model	Vocab.	#Param.	Size	RP	MRR@10	R@10(↑)	R@20	R@50	R@100
1	crossencoder-camembert-base-mmarcoFR	fr	110M	443MB	35.65	50.44	82.95	91.50	96.80	98.80
2	crossencoder-mMiniLMv2-L12-mmarcoFR	fr,99+	118M	471MB	34.37	51.01	82.23	90.60	96.45	98.40
3	crossencoder-mpnet-base-mmarcoFR	en	109M	438MB	29.68	46.13	80.45	87.90	93.15	96.60
4	crossencoder-distilcamembert-mmarcoFR	fr	68M	272MB	27.28	43.71	80.30	89.10	95.55	98.60
5	crossencoder-electra-base-french-mmarcoFR	fr	110M	443MB	28.32	45.28	79.22	87.15	93.15	95.75
6	crossencoder-mMiniLMv2-L6-mmarcoFR	fr,99+	107M	428MB	33.92	49.33	79.00	88.35	94.80	98.20

Training

Data

We use the French training samples from the mMARCO dataset, a multilingual machine-translated version of MS MARCO that contains 8.8M passages and 539K training queries. We sample 1M question-passage pairs from the official ~39.8M training triples with a positive-to-negative ratio of 4 (i.e., 25% of the pairs are relevant and 75% are irrelevant).

Implementation

The model is initialized from the cmarkea/distilcamembert-base checkpoint and optimized via the binary cross-entropy loss (as in monoBERT). It is fine-tuned on one 32GB NVIDIA V100 GPU for 10 epochs (i.e., 312.4k steps) using the AdamW optimizer with a batch size of 32, a peak learning rate of 2e-5 with warm up along the first 500 steps and linear scheduling. We set the maximum sequence length of the concatenated question-passage pairs to 512 tokens. We use the sigmoid function to get scores between 0 and 1.

Citation

@online{louis2023,
   author    = 'Antoine Louis',
   title     = 'crossencoder-distilcamembert-mmarcoFR: A Cross-Encoder Model Trained on 1M sentence pairs in French',
   publisher = 'Hugging Face',
   month     = 'september',
   year      = '2023',
   url       = 'https://huggingface.co/antoinelouis/crossencoder-distilcamembert-mmarcoFR',
}