ms-marco

Diploma thesis evaluating Retrieval-Augmented Generation (RAG) systems on MS MARCO and LegalBench-RAG using RAGChecker and RAGAs.

TypeScript client for the ltr-bert-sir ELSIE API (sir-elsie Space)

Python client for the ltr-bert-sir ELSIE API (sir-elsie Space)

Sign-coded late-interaction retrieval at 8 B/token: code, paper and full reproduction recipe for ReNeuIR @ SIGIR 2026

Multi-stage retrieval on MS MARCO (8.8M passages): BM25 + MonoT5 re-ranking + T5 query rewriting. NDCG@3 improved from 0.098 to 0.274.

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).

Local ONNX cross-encoder reranking for JavaScript/TypeScript. Zero API costs, framework-agnostic, with a Vercel-style API.

Fine-tuning cross-encoder re-rankers on MS MARCO, fusing their rankings, and expanding queries with an LLM, evaluated on TREC DL'19

Evaluating the effect of hard and soft watermarking on LLM text quality using centroid-based embedding analysis across 15,000 samples from the MS MARCO dataset. APPS597 Supervised Independent Study - University of Otago.

Small-scale replication of ColBERT late-interaction retrieval with BERT, MaxSim, MS MARCO training, indexing, and retrieval evaluation.

Multi-teacher LLM distillation for document reranking - three open-source teachers (Qwen2.5-14B-Instruct, DeepSeek-R1-Distill-Llama-8B, Mistral-Nemo-12B) teach a 7B student to rank and explain its rankings via QLoRA. Evaluated on TREC DL19/DL20.