FlowVec-v2

FlowVec-v2 is an 8-billion-parameter text-embedding model for Indian languages developed by Lingo Labs at IIT GN. It maps text in 21 Indic languages, across Bengali, Devanagari, Gujarati, Gurmukhi, Kannada, Malayalam, Odia, Tamil, Telugu, Perso-Arabic and Ol Chiki scripts, to 4096-dimensional vectors for retrieval, reranking, classification, clustering, bitext mining, semantic similarity and NLI.

Each task type comes with its own instruction, so one model serves search, cross-lingual matching and labelling without fine-tuning.

MTEB(Indic, v1) results

Scores from our own evaluation (two identical runs).

Task	FlowVec-v1	FlowVec-v2	Δ
BelebeleRetrieval	90.39	90.49	+0.10
BengaliSentimentAnalysis	92.0	92.06	+0.06
GujaratiNewsClassification	91.23	91.53	+0.30
HindiDiscourseClassification	43.59	43.85	+0.26
IN22ConvBitextMining	66.92	72.13	+5.21
IN22GenBitextMining	88.41	89.62	+1.21
IndicCrosslingualSTS	61.42	62.06	+0.64
MTOPIntentClassification	84.65	84.75	+0.10
MalayalamNewsClassification	91.8	92.75	+0.95
MultiHateClassification	67.38	67.19	-0.19
NepaliNewsClassification	97.01	97.01	+0.00
PunjabiNewsClassification	84.2	84.46	+0.26
SIB200ClusteringS2S	87.25	87.41	+0.16
SanskritShlokasClassification	64.9	71.46	+6.56
SentimentAnalysisHindi	76.02	76.23	+0.21
TweetSentimentClassification	46.56	46.91	+0.35
UrduRomanSentimentClassification	61.99	64.01	+2.02
WikipediaRerankingMultilingual	87.98	88.88	+0.90
XNLI	82.11	82.13	+0.02
XQuADRetrieval	94.35	94.36	+0.01
mean	78.01	78.96

Usage

from sentence_transformers import SentenceTransformer
import torch

model = SentenceTransformer("LingoIITGN/FlowVec-v2", model_kwargs={"torch_dtype": torch.bfloat16})
queries = model.encode(["भारत की राजधानी क्या है?"], prompt_name="XQuADRetrieval-query")
docs = model.encode(["नई दिल्ली भारत की राजधानी है।"])
print(model.similarity(queries, docs))

model.prompts lists every task instruction. For your own tasks, pass prompt="Instruct: <your task>\nQuery: ".