Embedding Model Training

Embedding Model Training

Apply the judgment of someone who has trained state-of-the-art retrievers and rerankers and shipped them into production search/RAG systems. The hard part is almost never the loss function — it is the data: which positives, and above all which negatives, you put in front of the model. Get hard-negative mining and false-negative removal right and a small model beats a big one.

How to use this skill

1. Read embedding-model-training-guide.md in this directory — the full reference (architectures, contrastive learning, hard-negative mining, training recipes, evaluation, production). Apply it.
2. For concrete artifacts to imitate — an InfoNCE bi-encoder training sketch, a positive-aware hard-negative-mining config, and a cross-encoder distillation outline — read examples.md.
3. Match the surrounding codebase/framework conventions ([[ml-frameworks]], [[training-frameworks]]); apply the data-quality and evaluation rules regardless.

The essentials (full rationale in the guide)

• Pick the right architecture for the latency budget. Bi-encoder (encode query and doc separately, compare with cosine/dot) is cheap and indexable — use it for first-stage retrieval over millions of docs. Cross-encoder (joint query+doc → score) is far more accurate but cannot be pre-indexed — use it as a reranker over the top-k. The default production shape is retrieve-then-rerank ([[rag-vector-databases]], [[recsys-ranking]]). If a reranker was the real need, do not train a bi-encoder.
• Contrastive learning is the workhorse. InfoNCE with in-batch negatives: every other doc in the batch is a negative for the current query. Large batches help because they give more negatives per step (DPR, E5). Tune temperature (often ~0.01–0.05 for normalized embeddings). Decide pooling (mean vs CLS/last-token) and L2-normalize when training with cosine.
• Hard negatives are the crux. Random/in-batch negatives are too easy and accuracy plateaus. Mine hard negatives with a retriever (top-ranked non-positives) — ANCE-style, refreshed periodically.
• The false-negative problem dominates web data. Naive top-k mining mislabels many true relevant passages as negatives. Positive-aware / threshold-based mining (NV-Retriever) anchors on the positive's score and drops candidates scoring too close to it → faster training and higher accuracy.
• Recipe: contrastive pretrain on weakly-supervised pairs → fine-tune on labeled + mined-hard-negative data → optionally distill from a cross-encoder teacher (soft scores / MarginMSE / KL). Instruction- prefixed embeddings let one model serve many tasks.
• Matryoshka (MRL) trains nested dimensions so you can truncate (e.g. 1024→256) at deploy time for cheaper storage/ANN with graceful accuracy loss. ColBERT multi-vector trades index size for accuracy via late interaction.
• Evaluate on MTEB (retrieval subset = BEIR) with nDCG@10 and recall@k — but watch contamination/overfitting and treat leaderboard ranks as time-scoped (verify against current docs).
• Production: the embedding dimension and distance metric must match the index exactly; quantize embeddings (int8/binary) to cut cost; and re-embed the whole corpus when the model version changes — query and corpus embeddings must come from the same model. Version your indexes.

Related skills

• [[rag-vector-databases]] — serving embeddings: ANN indexes (HNSW/IVF), hybrid search, RAG pipelines.
• [[recsys-ranking]] — two-tower retrieval and learning-to-rank; the same retrieve-then-rank shape.
• [[multimodal-ml]] — CLIP-style image/text and other multimodal embeddings (same contrastive core).
• [[fine-tuning-peft]] — LoRA/adapter fine-tuning of LLM-based embedding backbones.
• [[ml-frameworks]] / [[training-frameworks]] — PyTorch/JAX, FSDP/DeepSpeed, large-batch training.
• [[ml-evaluation-evals]] — building trustworthy eval sets and avoiding benchmark contamination.
• [[aiml-on-kubernetes]] — running the embedding/mining/eval jobs on GPUs at scale.