README.md

---
pipeline_tag: sentence-similarity
language: fr
license: mit
datasets:
- unicamp-dl/mmarco
metrics:
- recall
tags:
- passage-retrieval
library_name: transformers
base_model: almanach/camembert-base
model-index:
- name: spladev2-camembert-base-mmarcoFR
  results:
  - task:
      type: sentence-similarity
      name: Passage Retrieval
    dataset:
      type: unicamp-dl/mmarco
      name: mMARCO-fr
      config: french
      split: validation
    metrics:
    - type: recall_at_1000
      name: Recall@1000
      value: 89.86
    - type: recall_at_500
      name: Recall@500
      value: 85.96
    - type: recall_at_100
      name: Recall@100
      value: 73.94
    - type: recall_at_10
      name: Recall@10
      value: 46.33
    - type: map_at_10
      name: MAP@10
      value: 24.15
    - type: ndcg_at_10
      name: nDCG@10
      value: 29.58
    - type: mrr_at_10
      name: MRR@10
      value: 24.68
---

# spladev2-camembert-base-mmarcoFR

This is a [SPLADE-max](https://doi.org/10.48550/arXiv.2109.10086) model for **French** that can be used for semantic search. The model maps queries and passages to 
32k-dimensional sparse vectors which are used to compute relevance through cosine similarity.

## Usage

Start by installing the [library](https://huggingface.co/docs/transformers): `pip install -U transformers`. Then, you can use the model like this:

```python
import torch
from torch.nn.functional import relu, normalize
from transformers import AutoTokenizer, AutoModel

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/spladev2-camembert-base-mmarcoFR')
model = AutoModel.from_pretrained('antoinelouis/spladev2-camembert-base-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(**q_input)
    p_output = model(**p_input)

q_activations = torch.amax(torch.log1p(relu(q_output.logits * q_input['attention_mask'].unsqueeze(-1))), dim=1)
p_activations = torch.amax(torch.log1p(relu(p_output.logits * p_input['attention_mask'].unsqueeze(-1))), dim=1)

q_activations = normalize(q_activations, p=2, dim=1)
p_activations = normalize(p_activations, 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](https://ir-datasets.com/mmarco.html#mmarco/v2/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*](https://huggingface.co/spaces/antoinelouis/decouvrir) leaderboard.

## Training

#### Data

The model is trained on the French training samples of the [mMARCO](https://huggingface.co/datasets/unicamp-dl/mmarco) dataset, a multilingual machine-translated version of MS MARCO that 
contains 8.8M passages and 539K training queries. We sample 12.8M (q, p+, p-) triples from the official ~39.8M [training triples](https://microsoft.github.io/msmarco/Datasets.html#passage-ranking-dataset) 
with BM25 negatives.

#### Implementation

The model is initialized from the [almanach/camembert-base](https://huggingface.co/almanach/camembert-base) checkpoint and optimized via a combination of the InfoNCE 
ranking loss with a temperature of 0.05 and the FLOPS regularization loss with quadratic increase of lambda until step 33k after which it remains constant with lambda_q=3e-4 
and lambda_d=1e-4. The model is fine-tuned on one 80GB NVIDIA H100 GPU for 100k steps using the AdamW optimizer with a batch size of 128, a peak learning rate 
of 2e-5 with warm up along the first 4000 steps and linear scheduling. The maximum sequence lengths for questions and passages length were fixed to 32 and 128 tokens. 
Relevance scores are computed with the cosine similarity.

## Citation

```bibtex
@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',
}
```