eventpulse-bge-m3

eventpulse-bge-m3

A fine-tuned version of BAAI/bge-m3 for semantic matching of user search queries against event poster titles (concerts, films, theatre).

Trained for EventPulse_3.0 — a multilingual event discovery and aggregation application.

What it does

Maps user queries to event titles in embedding space so that semantically related pairs have high cosine similarity.

The model handles real-world search noise:

• Typos — "metalica" → "Metallica"
• Transliteration — "колдплей" → "Coldplay", "korj" → "Макс Корж"
• Abbreviations — "дюна 2" → "Дюна: Часть вторая"
• Cross-language matching — "john wick 4" → "Джон Уик 4"
• Partial titles — "кастанеда" → "KASTANEDA: ARENA ŠOU | Žalgirio arena"

Languages

English, Russian, Lithuanian, Polish — and mixed-language inputs.

Training data

~7 400 triplets (query, positive, negative) built from real event poster titles scraped from:

• biletai.lt — concerts in Vilnius
• concertful.pl — concerts in Warsaw
• afisha.me / relax.by — films in Minsk

Queries were generated via programmatic augmentation (typos, transliteration, truncation) and LLM-assisted generation (Gemini API).

Training

• Loss: MultipleNegativesRankingLoss (sentence-transformers)
• Base model: BAAI/bge-m3
• Epochs: 1
• Batch size: 32
• Learning rate: 2e-5

Usage

from sentence_transformers import SentenceTransformer
import numpy as np

model = SentenceTransformer("OlegChausov/eventpulse-bge-m3")

query = "металика"
title = "Metallica"

q_emb = model.encode(query, normalize_embeddings=True)
t_emb = model.encode(title, normalize_embeddings=True)

score = float(np.dot(q_emb, t_emb))
print(f"Similarity: {score:.3f}")  # ~0.85

Example scores