cmarkea/bloomz-3b-retriever

Note

We now strongly recommend using the Bloomz-3b-retriever-v2 model, which offers significantly superior performance.

Bloomz-3b-retriever

We introduce the Bloomz-3b-retriever based on the Bloomz-3b-sft-chat model. This model enables the creation of an embedding representation of text and queries for a retrieval task, linking queries to documents. The model is designed to be cross-language, meaning it is language-agnostic (English/French). This model is ideal for Open Domain Question Answering (ODQA), projecting queries and text with an algebraic structure to bring them closer together.

Training

It is a bi-encoder trained on a corpus of context/query pairs, with 50% in English and 50% in French. The language distribution for queries and contexts is evenly split (1/4 French-French, 1/4 French-English, 1/4 English-French, 1/4 English-English). The learning objective is to bring the embedding representation of queries and associated contexts closer using a contrastive method. The loss function is defined in Deep Metric Learning using Triplet Network.

Benchmark

Based on the SQuAD evaluation dataset (comprising 6000 queries distributed over 1200 contexts grouped into 35 themes), we compare the performance in terms of the average top contexter value for a query (Top-mean), the standard deviation of the average top (Top-std), and the percentage of correct queries within the top-1, top-5, and top-10. We compare the model with a TF-IDF trained on the SQuAD train sub-dataset (we want a fixed algebraic structure for the vector database instead of a variable structure every time we add a new document, then the IDF part has frozen), CamemBERT, Sentence-BERT, and finally our model. We observe these performances in both monolingual and cross-language contexts (query in French and context in English).

Model (FR/FR)	Top-mean	Top-std	Top-1 (%)	Top-5 (%)	Top-10 (%)
TF-IDF	128	269	23	46	56
CamemBERT	417	347	1	2	3
Sentence-BERT	11	41	43	71	82
Bloomz-560m-retriever	10	47	51	78	86
Bloomz-3b-retriever	9	37	50	79	87

Model (EN/FR)	Top-mean	Top-std	Top-1 (%)	Top-5 (%)	Top-10 (%)
TF-IDF	607	334	0	0	0
CamemBERT	432	345	0	1	1
Sentence-BERT	12	47	44	73	83
Bloomz-560m-retriever	10	44	49	77	86
Bloomz-3b-retriever	9	38	50	78	87

We observed that TF-IDF loses robustness in cross-language scenarios (even showing lower performance than CamemBERT, which is a model specialized in French). This can be explained by the fact that a Bag-Of-Words method cannot support this type of issue because, for a given sentence between two languages, the embedding vectors will be significantly different.

CamemBERT exhibits poor performance, not because it poorly groups contexts and queries by themes, but because a meta-cluster appears, separating contexts and queries (as illustrated in the image below), making this type of modeling inappropriate in a retriever context.

How to Use Bloomz-3b-retriever

The following example utilizes the API Pipeline of the Transformers library.

import numpy as np
from transformers import pipeline
from scipy.spatial.distance import cdist

retriever = pipeline('feature-extraction', 'cmarkea/bloomz-3b-retriever')

# Inportant: take only last token!
infer = lambda x: [ii[0][-1] for ii in retriever(x)]

list_of_contexts = [...]
emb_contexts = np.concatenate(infer(list_of_contexts), axis=0)
list_of_queries = [...]
emb_queries = np.concatenate(infer(list_of_queries), axis=0)

# Important: take l2 distance!
dist = cdist(emb_queries, emb_contexts, 'euclidean') 
top_k = lambda x: [
    [list_of_contexts[qq] for qq in ii]
    for ii in dist.argsort(axis=-1)[:,:x]
]

# top 5 nearest contexts for each queries
top_contexts = top_k(5)

Citation

@online{DeBloomzRet,
  AUTHOR = {Cyrile Delestre},
  ORGANIZATION = {Cr{\'e}dit Mutuel Ark{\'e}a},
  URL = {https://huggingface.co/cmarkea/bloomz-3b-retriever},
  YEAR = {2023},
  KEYWORDS = {NLP ; Transformers ; LLM ; Bloomz},
}