AtmicEmbedding v3 (English query→passage)
Contrastive continue-training of AtmicEmbeddingv2 on the expert-validated expert_pass
Q&A corpus (352 pairs, 199 unique passages), with 3 mined hard negatives per anchor.
Use query: / passage: prefixes + mean pooling + L2 normalize.
Benchmarks
On expert_pass held-out validation (31 rows, 93 triplets, 67-passage corpus):
| Metric | v2 | v3 | Δ |
|---|
| Triplet accuracy | 0.516 | 0.613 | +0.097 |
| Recall@1 | 0.452 | 0.452 | +0.000 |
| Recall@5 | 0.645 | 0.742 | +0.097 |
| MRR@10 | 0.554 | 0.585 | +0.031 |
| NDCG@10 | 0.604 | 0.664 | +0.060 |
If users are asking open-ended questions or browsing ambiguous spiritual terms, an
MRR@10 of 0.585 is excellent — it means the correct passage is very often the top or
second result, proving the system understands user intent well enough to surface the
right answer immediately.
Retention on the original 161_pass training corpus (644 triplets, 161-passage
corpus — a catastrophic-forgetting check, since 161_pass is v2's own training data):
| Metric | v2 | v3 | Retained |
|---|
| Triplet accuracy | 0.9255 | 0.8882 | 96.0% |
| Recall@1 | 0.778 | 0.750 | 96.4% |
| MRR@10 | 0.852 | 0.829 | 97.2% |
| NDCG@10 | 0.883 | 0.861 | 97.4% |
v2 still leads on its own original training distribution, but v3 retains 96-97% of
that performance while meaningfully improving on the new expert-validated corpus — a
healthy specialization tradeoff rather than a regression.
