minicoil-training
miniCOIL v1 sparse neural retrieval model training methodology: per-word linear layer training with triplet loss, word vocabulary from frequency-filtered English words, sparse retrieval with jina-embeddings-v2-small-en encoder, sparse embedding training pipelines, OpenWebText sentence extraction for self-supervised training, semi-hard triplet mining, and evaluation on BEIR benchmarks. English-only miniCOIL v1 (jina-embeddings, 30k words, mxbai-embed-large-v1 mining). Reference for building, training, and deploying sparse neural retrieval models compatible with inverted indexes and Qdrant.
npx skills add majiayu000/claude-skill-registry --skill minicoil-training --agent claude-code
Same command for any agent — swap --agent for codex, cursor, copilot.
Weekly change comes from our own snapshots, not the repository page — it measures attention, not adoption.
# miniCOIL Training Reference Complete reference for training miniCOIL sparse neural retrieval models. Covers the full pipeline from word vocabulary construction through evaluation. ## What is miniCOIL? miniCOIL is a lightweight sparse neural retrieval model created by Qdrant. It adds contextual semantic awareness to BM25-style keyword matching while preserving inverted-index compatibility. **The core problem:** The same word has different meanings in different contexts ("fruit bat" vs. "baseball bat"). BM25 is blind to this. miniCOIL trains a tiny per-word linear layer that projects sentence embeddings into a 4D meaning space, letting the inverted index distinguish word senses. ### Scoring Formula ``` score(D, Q) = SUM_{i: q_i in D} IDF(q_i) * dot(layer_i(emb_Q), layer_i(emb_D)) ``` Where `layer_i` is the per-word Linear(512, 4) + tanh projection for word `i`. | Component | Source | What it captures | |---------------------|----------|--------------------------------------------------------| | IDF(q_i) | BM25 | Inverse document frequency — rare terms matter more | | layer_i(emb_Q) | miniCOIL | 4D meaning vector for word i in query context | | layer_i(emb_D) | miniCOIL | 4D meaning
- What is miniCOIL?
- Scoring Formula
- Comparison with Alternatives
- Architecture
- Key Design Decisions
- Full Training Pipeline
- Step 1: Build Word Vocabulary
- Step 2: Extract Training Sentences
- Step 3: Per-Word Training
- Step 4: Sparse Encoding
- Step 5: Evaluation
- Triplet Mining Algorithm
- For each triplet to mine:
- Adaptive triplet count
English monolingual (BEIR) minicoil evaluate --datasets nq quora fiqa hotpotqa msmarco
What does the minicoil-training skill do?
miniCOIL v1 sparse neural retrieval model training methodology: per-word linear layer training with triplet loss, word vocabulary from frequency-filtered English words, sparse retrieval with jina-embeddings-v2-small-en encoder, sparse embedding training pipelines, OpenWebText sentence extraction for self-supervised training, semi-hard triplet mining, and evaluation on BEIR benchmarks. English-only miniCOIL v1 (jina-embeddings, 30k words, mxbai-embed-large-v1 mining). Reference for building, training, and deploying sparse neural retrieval models compatible with inverted indexes and Qdrant.
How do I install it?
Run `npx skills add majiayu000/claude-skill-registry --skill minicoil-training --agent claude-code` — it drops the skill into your project so the agent can pick it up. Swap the --agent value for codex, cursor or copilot if you use one of those.
Where does this skill come from?
From majiayu000/claude-skill-registry, a repository with 534 stars. We read it straight from the repository tree rather than a submitted listing, so what you see here is what is actually published.
Is a popular skill a good skill?
Not necessarily. Stars measure attention, not adoption — a repository can trend for a week and be abandoned. That is why we show the weekly change from our own snapshots next to the total, instead of a single flattering number.