AurelPx/e5-multilingual-frenchboost-v1

E5 Multilingual FrenchBoost v1

E5 Multilingual FrenchBoost v1 is a multilingual sentence embedding model boosted for French, derived from intfloat/multilingual-e5-base.

This is the first public release (model-v1). The goal is practical French semantic search, reranking, clustering, and similarity while preserving the broad multilingual E5 embedding space as much as possible.

Intended use

Use this model for French-heavy multilingual embeddings in retrieval, reranking, semantic search, clustering, duplicate detection, and similarity scoring. Like E5, it works best with explicit prefixes:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("AurelPx/e5-multilingual-frenchboost-v1")
queries = ["query: Quelle est la capitale de la France ?"]
passages = ["passage: Paris est la capitale de la France."]

query_embeddings = model.encode(queries, normalize_embeddings=True)
passage_embeddings = model.encode(passages, normalize_embeddings=True)
scores = query_embeddings @ passage_embeddings.T

Training summary

• Base model: intfloat/multilingual-e5-base
• Method: conservative LoRA fine-tuning followed by merge for easy Sentence Transformers inference
• Data mixture: French retrieval/reranking signals, MIRACL-style hard-negative ranking, French STS/paraphrase/NLI signals, multilingual replay, and baseline embedding distillation
• Selection: held-out French retrieval + STS proxy evaluation with anti-regression guardrails and no train/eval leakage
• Export checks: non-zero LoRA weights, adapter reload parity, merged-model embedding parity

Evaluation

Internal evaluation was run with a memory-safe MTEB(fra)-safe v2 protocol: one task per isolated process, using bounded variants for the historically crash-prone clustering tasks. These numbers should be read as a transparent development benchmark, not as a final SOTA claim.

• Mean score over 25 tasks: 0.5636
• Benchmark label: MTEB(fra)-safe v2 GPU-fp16 conservative-lora

Category scores

category	score	n_tasks
Classification	0.6384	6
STS	0.7423	3
Reranking	0.7681	2
PairClass.	0.5538	1
Clustering	0.4402	7
Retrieval	0.5062	5
Summarization	0.3294	1

Per-task comparison

The table below compares this V1 checkpoint against the previous internal LoRA candidate when that reference run is available locally.

task	v1_score	reference_score	delta
SyntecRetrieval	0.8231	0.3292	0.4939
BSARDRetrieval	0.5856	0.1622	0.4234
SyntecReranking	0.8525	0.5442	0.3083
MasakhaNEWSClusteringS2S	0.5543	0.2479	0.3064
AlloprofRetrieval	0.3558	0.0557	0.3001
AlloprofReranking	0.6837	0.5077	0.1760
AlloProfClusteringP2P	0.6263	0.4635	0.1628
XPQARetrieval	0.4571	0.3401	0.1170
SummEvalFr	0.3294	0.2921	0.0373
STS22	0.6485	0.6187	0.0298
MTOPDomainClassification	0.8266	0.8025	0.0241
MassiveScenarioClassification	0.6710	0.6482	0.0228

Current limitations

This is a V1. It is already stronger in reranking in the internal evaluation, but the model card will be updated as broader validation comes in. The next version will focus on better robustness across retrieval, clustering, and pair classification while keeping STS gains.

Citation / attribution

This model builds on multilingual E5 by Microsoft / intfloat and uses the Sentence Transformers ecosystem.