Sentence-Transformers Information Retrieval example on Chinese

Code and created datasets for our ACL 2022 paper: "Contextual Fine-to-Coarse Distillation for Coarse-grained Response Selection in Open-Domain Conversations"

ViIR: The Unified Framework for Fine-tuning Vietnamese Information Retrieval Models with Various Tuning Statergies.

Official codebase for the ACL 2026 MeLLM Workshop paper "The Multilingual Curse at the Retrieval Layer: Evidence from Amharic"

High-accuracy job classification system using Sentence Transformers. Maps job titles & descriptions to 1,016 O*NET-SOC categories. 100% Top-1 accuracy on real job postings. Fast CPU inference (<100ms). 126K+ training samples from 8 O*NET data sources.

Two-stage retrieve-and-rank neural product search on Amazon ESCI: a dense bi-encoder retriever with hard-negative mining + a DeBERTa cross-encoder reranker over Exact/Substitute/Complement/Irrelevant labels. NDCG@10 0.71 (+16% vs BM25), 0.74 micro-F1.

Neural search engine with bi-encoder + cross-encoder re-ranking, BM25 hybrid search, query expansion, typo correction and multi-language support — processes 10M queries/day

Baseline models for searching for movie plots from Wikipedia articles. Techniques include BM25 (lexical search), bi/cross-encoding (semantic search), and retrieval-augmented generation (RAG) using Mistal 7B through Fireworks.ai.

Proof of concept for large language model summarization of medical journal articles for different reading levels

Comparative study of parameter-efficient fine-tuning (PEFT) strategies for biomedical NER on top of GLiNER — including soft prompt tuning, embedding injection, and a custom in-place embedding extension that matches full fine-tuning performance at 13% of trainable parameters.

Évaluation de la pertinence (question ↔ article juridique) en français. Pipeline complet (prépa → modèles → soumission) avec CamemBERT en bi-encodeur calibré (MSE/Spearman), + variantes cross-encoder.

Powered by a catalog of 190+ products, this engine delivers high-precision results using semantic embeddings and vector similarity principles. By mapping product data into high-dimensional space and calculating the cosine similarity between search queries and items, it identifies matches based on intent and meaning rather than just keywords.

Recommendation systems overview and an MLOps TFX-pipeline implementation

InsureLLM RAG Challenge — Two-stage retrieval pipeline (Bi-Encoder + Cross-Encoder) with context compression

Tiny semantic search engine with bi-encoder retrieval and cross-encoder reranking — built to understand how production search works under the hood.

Exploring fast & accurate zero-shot text classification

Controlled depth ablation of a BERT bi-encoder across training budgets and seeds on three BEIR tasks (nfcorpus, scifact, fiqa). L3–L12 is flat within seed noise at 20K steps; 80K training degrades every depth on zero-shot transfer (−45% NDCG@10 on fiqa for L12).