History — 2026 week 42 (Oct 12 – Oct 18, 2026)
Newest first.
2026-10-18 — Fine-tuning embeddings: gaps, data, losses, hyperparameters (4 new, 8 updated)
A general-purpose embedding model knows the words in a legal or clinical corpus but not what the field treats as equivalent, and fine-tuning is how that gap closes. Domain-Specific Semantic Gaps covers the problem; Dataset Types for Embedding Fine-Tuning the six shapes training data comes in, from positive pairs to hard-negative sets, with suggested minimum sizes; Embedding Loss Functions which loss each shape calls for, with MultipleNegativesRankingLoss as the default; and Hyperparameter Tuning for Embedding Fine-Tuning how to set learning rate, batch size, warmup and epochs from the loss curve. Drawn from Ritik Jain’s member-only guide, whose reported retrieval gains are the author’s own.
Corrections
- The Complete Guide to Fine-Tuning Embedding Models linked to a ritikjain51.medium.com address; it now links to the post on rittikjain.medium.com and is marked member-only.
Concepts — Domain-Specific Semantic Gaps · Dataset Types for Embedding Fine-Tuning · Embedding Loss Functions · Hyperparameter Tuning for Embedding Fine-Tuning Updated — The Complete Guide to Fine-Tuning Embedding Models (rewritten as key ideas) · Ritik Jain · Embedding Fine-tuning · Contrastive Learning · Hard Negative Mining · Sentence Transformers · Synonyms and Vocabulary Management · Model Selection and Fine-Tuning Evaluation
2026-10-18 — sqlite-sparse: learned sparse retrieval inside SQLite (1 new, 3 updated)
Learned Sparse Retrieval usually means a search cluster, but sqlite-sparse runs it inside a single SQLite file with no model on the query path. Documents are encoded at insert with an inference-free OpenSearch sparse encoder, and a query is scored exactly over posting lists stored as database rows, using only a per-token weight table. The note covers how it keeps the SPLADE-style vocabulary head that llama.cpp discards. It also covers the costs: indexing is far slower, and the index larger, than SQLite’s built-in BM25 search. Its latency and quality figures are the project’s own, on one million MS MARCO passages and four BEIR subsets, and have not been replicated.
Tools — sqlite-sparse (SQLite extension, no query-time model) Updated — Learned Sparse Retrieval · Search Platforms · psql_bm25s
2026-10-13 — Catastrophic forgetting and single- vs multi-domain fine-tuning (1 new, 5 updated)
A retriever fine-tuned on one catalog gets better there and worse everywhere else: in Thierry Damiba’s SPLADE experiments, a model trained on Amazon data lost to the off-the-shelf model on Home Depot and fell well below BM25 on MS MARCO. Catastrophic Forgetting now covers the three responses: single-domain training for one catalog, multi-domain training that gives up some peak accuracy for consistency, and per-domain LoRA experts with query routing from RouterRetriever. It also covers how to catch forgetting with a held-out general set at every checkpoint. Drawn from Damiba’s series and Evgeniya Sukhodolskaya’s talk on the same work.
Concepts — Catastrophic Forgetting (also: multi-domain fine-tuning) Updated — Embedding Fine-tuning · SPLADE · LoRA · Model Selection and Fine-Tuning Evaluation · Fine-Tuning Sparse Embeddings for E-Commerce Search
2026-10-13 — Benchmark contamination and inflated retrieval leaderboards (1 new, 5 updated)
When a public benchmark’s corpus, queries and labels leak into training data, a model can score well by remembering the answers, and the score stops predicting performance on unseen data. Benchmark Contamination now covers how that happens, with MS MARCO underneath nearly every retrieval model and BEIR’s public corpora a standing caveat, why the leaderboard cannot tell memorization from skill, and how RTEB detects it by comparing open and privately held datasets. It also covers the 2026 temporary removal of RTEB’s private column over vendor-access concerns: hiding the test set moves the incentive problem to whoever holds the secret. MTEB and Retrieval Benchmarks and Leaderboards now link to it.
Concepts — Benchmark Contamination (test data leaking into training data) Updated — RTEB · MTEB · BEIR · MS MARCO · Retrieval Benchmarks and Leaderboards
2026-10-12 — Decay and forgetting in agentic memory (0 new, 2 updated)
A memory store that never forgets keeps applying a preference from a life stage the user has already left. Agentic Memory now covers forgetting at three granularities: a blanket TTL, memory that lives only as long as one conversation as in django-dice, and per-fact decay scores in DICE, where stable preferences age slowly and transient states fast. It also covers how decay pushes an ageing fact out of a capped context window before anything deletes it, how contradicting evidence lowers a fact’s confidence without overwriting it, and why forgetting should leave an audit trail. Drawn from the DICE and Embabel DICE material and Hajer Bouafif’s OpenSearch guidance.
Updated — Agentic Memory (new Decay and Forgetting section) · Agentic Memory for Search Personalization (staleness failure mode links to it)