| # Vinci v1 Technical Report Draft |
|
|
| ## Summary |
|
|
| Vinci is a DB-native embedding model for Cilow's Lattice retrieval pipeline. The |
| v1 research target is public MTEB Retrieval quality with a BGE-M3 student, |
| Matryoshka dimensions `1024` and `256`, and commercial-safe teachers and |
| training sources. |
|
|
| ## Model |
|
|
| - Canonical model id: `Cilow/Vinci` |
| - Student base: `BAAI/bge-m3` |
| - Model family: `vinci` |
| - Pipeline: `Lattice` |
| - Similarity: cosine |
| - Public claim rule: 2x means error reduction, where `error = 1 - metric`. |
|
|
| ## Training Objective |
|
|
| The planned objective combines Matryoshka contrastive retrieval and teacher |
| margin distillation: |
|
|
| `L = sum_m w_m * L_InfoNCE(m) + lambda_margin * L_teacher_margin + lambda_struct * L_structure` |
|
|
| where `m` is one of `1024` or `256`. Cilow truth labels override generic teacher |
| similarity for stale, superseded, contradicted, or numerically wrong facts. |
|
|
| ## MTEB Results |
|
|
| | Model | Track | Task | nDCG@10 | Recall@10 | MRR@10 | |
| |---|---|---|---:|---:|---:| |
| | Cilow/Vinci | embedding | SciFact | 0.7100 | 0.7900 | 0.6100 | |
| | Cilow/Vinci | embedding | FiQA2018 | 0.7600 | 0.8400 | 0.6600 | |
| | Cilow/Vinci | embedding | NFCorpus | 0.6600 | 0.7400 | 0.5600 | |
|
|
| ## 2x Scorecard |
|
|
| | Task | Metric | Baseline | Candidate | Error Reduction | 2x | |
| |---|---|---:|---:|---:|---| |
| | FiQA2018 | ndcg_at_10 | 0.7700 | 0.7600 | 0.9583 | False | |
| | NFCorpus | ndcg_at_10 | 0.6700 | 0.6600 | 0.9706 | False | |
| | SciFact | ndcg_at_10 | 0.7200 | 0.7100 | 0.9655 | False | |
|
|
| ## Leakage Check |
|
|
| Passed: `True`. Train hashes: `5`. Eval hashes: `2`. Overlap: `0`. |
|
|
| ## Publication Status |
|
|
| This report is a scaffold until full MTEB Retrieval runs, local SentenceTransformer |
| eval, TEI eval, and Lattice+reranker eval are attached. Model-only and full-pipeline |
| claims must remain separate. |
|
|
