đ ïž Usage | đ Evaluation | đ€ Training | đ Citation
This is a [sentence-transformers](https://www.SBERT.net) model. It maps questions and paragraphs 768-dimensional dense vectors and should be used for semantic search. The model uses an [CamemBERT-L8](https://huggingface.co/antoinelouis/camembert-L8) backbone, which is a pruned version of the pre-trained [CamemBERT](https://huggingface.co/camembert-base) checkpoint with 26% less parameters, obtained by [dropping the top-layers](https://doi.org/10.48550/arXiv.2004.03844) from the original model. The model was trained on the **French** portion of the [mMARCO](https://huggingface.co/datasets/unicamp-dl/mmarco) retrieval dataset. ## Usage Here are some examples for using this model with [Sentence-Transformers](#using-sentence-transformers), [FlagEmbedding](#using-flagembedding), or [Huggingface Transformers](#using-huggingface-transformers). #### Using Sentence-Transformers Start by installing the [library](https://www.SBERT.net): `pip install -U sentence-transformers`. Then, you can use the model like this: ```python from sentence_transformers import SentenceTransformer queries = ["Ceci est un exemple de requĂȘte.", "Voici un second exemple."] passages = ["Ceci est un exemple de passage.", "Et voilĂ un deuxiĂšme exemple."] model = SentenceTransformer('antoinelouis/biencoder-camembert-L8-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](https://github.com/FlagOpen/FlagEmbedding/): `pip install -U FlagEmbedding`. Then, you can use the model like this: ```python from FlagEmbedding import FlagModel queries = ["Ceci est un exemple de requĂȘte.", "Voici un second exemple."] passages = ["Ceci est un exemple de passage.", "Et voilĂ un deuxiĂšme exemple."] model = FlagModel('antoinelouis/biencoder-camembert-L8-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](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 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 requĂȘte.", "Voici un second exemple."] passages = ["Ceci est un exemple de passage.", "Et voilĂ un deuxiĂšme exemple."] tokenizer = AutoTokenizer.from_pretrained('antoinelouis/biencoder-camembert-L8-mmarcoFR') model = AutoModel.from_pretrained('antoinelouis/biencoder-camembert-L8-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 We evaluate the model 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. Below, we compare the model performance with other CamemBERT-based biencoder models fine-tuned on the same dataset. We report the mean reciprocal rank (MRR), normalized discounted cumulative gainand (NDCG), mean average precision (MAP), and recall at various cut-offs (R@k). | | model | #Param. | Size | R@500 | R@100(â) | R@10 | MRR@10 | NDCG@10 | MAP@10 | |---:|:-------------------------------------------------------------------------------------------------------------|--------:|------:|-------:|---------:|-------:|-------:|--------:|-------:| | 1 | [biencoder-camembert-base-mmarcoFR](https://huggingface.co/antoinelouis/biencoder-camembert-base-mmarcoFR) | 111M | 445MB | 89.1 | 77.8 | 51.5 | 28.5 | 33.7 | 27.9 | | 2 | [biencoder-camembert-L10-mmarcoFR](https://huggingface.co/antoinelouis/biencoder-camembert-L10-mmarcoFR) | 96M | 386MB | 87.8 | 76.7 | 49.5 | 27.5 | 32.5 | 27.0 | | 3 | **biencoder-camembert-L8-mmarcoFR** | 82M | 329MB | 87.4 | 75.9 | 48.9 | 26.7 | 31.8 | 26.2 | | 4 | [biencoder-camembert-L6-mmarcoFR](https://huggingface.co/antoinelouis/biencoder-camembert-L6-mmarcoFR) | 68M | 272MB | 86.7 | 74.9 | 46.7 | 25.7 | 30.4 | 25.1 | | 5 | [biencoder-camembert-L4-mmarcoFR](https://huggingface.co/antoinelouis/biencoder-camembert-L4-mmarcoFR) | 54M | 216MB | 85.4 | 72.1 | 44.2 | 23.7 | 28.3 | 23.2 | | 6 | [biencoder-camembert-L2-mmarcoFR](https://huggingface.co/antoinelouis/biencoder-camembert-L2-mmarcoFR) | 40M | 159MB | 81.0 | 66.3 | 38.5 | 20.1 | 24.3 | 19.7 | *** ## Training #### Data We use the French training samples from 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 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](https://huggingface.co/datasets/sentence-transformers/msmarco-hard-negatives) distillation dataset. #### Implementation The model is initialized from the [camembert-L8](https://huggingface.co/antoinelouis/camembert-L8) checkpoint and optimized via the cross-entropy loss (as in [DPR](https://doi.org/10.48550/arXiv.2004.04906)) with a temperature of 0.05. It is fine-tuned on one 32GB NVIDIA V100 GPU for 39k steps (or 40 epochs) using the AdamW optimizer with a batch size of 512, a peak learning rate of 2e-5 with warm up along the first 3900 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 ```bibtex @online{louis2023, author = 'Antoine Louis', title = 'biencoder-camembert-L8-mmarcoFR: A Biencoder Model Trained on French mMARCO', publisher = 'Hugging Face', month = 'may', year = '2023', url = 'https://huggingface.co/antoinelouis/biencoder-camembert-L8-mmarcoFR', } ```