biencoder-camembert-L4-mmarcoFR

This is a lightweight dense single-vector bi-encoder model for French that can be used for semantic search. The model maps queries and passages to 768-dimensional dense vectors which are used to compute relevance through cosine similarity. It uses a CamemBERT-L4 backbone, which is a pruned version of the pre-trained CamemBERT checkpoint with 51% less parameters, obtained by dropping the top-layers from the original model.

Usage

Here are some examples for using this 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 SentenceTransformer

queries = ["Ceci est un exemple de requête.", "Voici un second exemple."]
passages = ["Ceci est un exemple de passage.", "Et voilà un deuxième exemple."]

model = SentenceTransformer('antoinelouis/biencoder-camembert-L4-mmarcoFR')

q_embeddings = model.encode(queries, normalize_embeddings=True)
p_embeddings = model.encode(passages, normalize_embeddings=True)

similarity = q_embeddings @ p_embeddings.T
print(similarity)

Using FlagEmbedding

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

from FlagEmbedding import FlagModel

queries = ["Ceci est un exemple de requête.", "Voici un second exemple."]
passages = ["Ceci est un exemple de passage.", "Et voilà un deuxième exemple."]

model = FlagModel('antoinelouis/biencoder-camembert-L4-mmarcoFR')

q_embeddings = model.encode(queries, normalize_embeddings=True)
p_embeddings = model.encode(passages, normalize_embeddings=True)

similarity = q_embeddings @ p_embeddings.T
print(similarity)

Using Transformers

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

import torch
from torch.nn.functional import normalize
from transformers import AutoTokenizer, AutoModel

def mean_pooling(model_output, attention_mask):
    """ Perform mean pooling on-top of the contextualized word embeddings, while ignoring mask tokens in the mean computation."""
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)


queries = ["Ceci est un exemple de requête.", "Voici un second exemple."]
passages = ["Ceci est un exemple de passage.", "Et voilà un deuxième exemple."]

tokenizer = AutoTokenizer.from_pretrained('antoinelouis/biencoder-camembert-L4-mmarcoFR')
model = AutoModel.from_pretrained('antoinelouis/biencoder-camembert-L4-mmarcoFR')

q_input = tokenizer(queries, padding=True, truncation=True, return_tensors='pt')
p_input = tokenizer(passages, padding=True, truncation=True, return_tensors='pt')
with torch.no_grad():
    q_output = model(**encoded_queries)
    p_output = model(**encoded_passages)
q_embeddings = mean_pooling(q_output, q_input['attention_mask'])
q_embedddings = normalize(q_embeddings, p=2, dim=1)
p_embeddings = mean_pooling(p_output, p_input['attention_mask'])
p_embedddings = normalize(p_embeddings, p=2, dim=1)

similarity = q_embeddings @ p_embeddings.T
print(similarity)

Evaluation

The model is evaluated on the smaller development set of mMARCO-fr, which consists of 6,980 queries for a corpus of 8.8M candidate passages. We report the mean reciprocal rank (MRR), normalized discounted cumulative gainand (NDCG), mean average precision (MAP), and recall at various cut-offs (R@k). To see how it compares to other neural retrievers in French, check out the DécouvrIR leaderboard.

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 do not employ the BM25 negatives provided by the official dataset but instead sample harder negatives mined from 12 distinct dense retrievers, using the msmarco-hard-negatives distillation dataset.

Implementation

The model is initialized from the camembert-L4 checkpoint and optimized via the cross-entropy loss (as in DPR) with a temperature of 0.05. It is fine-tuned on one 32GB NVIDIA V100 GPU for 17.4k steps (or 40 epochs) using the AdamW optimizer with a batch size of 1152, a peak learning rate of 2e-5 with warm up along the first 1736 steps and linear scheduling. We set the maximum sequence lengths for both the questions and passages to 128 tokens. We use the cosine similarity to compute relevance scores.

Citation

@online{louis2024decouvrir,
	author    = 'Antoine Louis',
	title     = 'DécouvrIR: A Benchmark for Evaluating the Robustness of Information Retrieval Models in French',
	publisher = 'Hugging Face',
	month     = 'mar',
	year      = '2024',
	url       = 'https://huggingface.co/spaces/antoinelouis/decouvrir',
}