AtmicEmbedding v3 (English query→passage)
Contrastive continue-training of AtmicEmbeddingv2 on a much larger combined corpus: 452 expert Q&A pairs + 1,356 Claude-generated paraphrase/scenario questions (same answers)
- 1,368 pre-mined theoretical-dialogue triplets (question/positive/negative_1-3, deduped
where slots repeated — 3,176 total training rows, 1,666 unique passages). Every row gets
3 negatives; where a source doesn't already supply them, negative_1 is drawn from a
100-chunk pool of genuinely mistaken interpretations of Ramana's teaching (extracted from
a scholarly critique, then rewritten to strip named-interpreter attribution), and all
negative selection uses hybrid dense+BM25 scoring (Reciprocal Rank Fusion).
Use
query:/passage:prefixes + mean pooling + L2 normalize.
Benchmarks
Three-way comparison — base intfloat/multilingual-e5-large (zero domain fine-tuning) vs.
v2 vs. v3 — on two test sets.
Own held-out test data (275 rows, 581-passage corpus):
| Metric | base e5-large | v2 | v3 |
|---|---|---|---|
| Triplet accuracy | 0.611 | 0.532 | 0.884 |
| Recall@1 | 0.251 | 0.233 | 0.415 |
| Recall@5 | 0.491 | 0.466 | 0.651 |
| MRR@10 | 0.361 | 0.348 | 0.521 |
| NDCG@10 | 0.399 | 0.392 | 0.566 |
v3 beats both base and v2 by a wide margin here — the larger, more diverse training set (theoretical dialogue data plus paraphrase/scenario questions) generalizes well beyond either the un-fine-tuned base model or v2's narrower specialization.
Full 161_pass dataset (644 triplets — v2's OWN original training data, used as a retention / catastrophic-forgetting check):
| Metric | base e5-large | v2 | v3 |
|---|---|---|---|
| Triplet accuracy | 0.775 | 0.926 | 0.812 |
| Recall@1 | 0.592 | 0.778 | 0.571 |
| Recall@5 | 0.800 | 0.957 | 0.789 |
| MRR@10 | 0.682 | 0.852 | 0.675 |
| NDCG@10 | 0.717 | 0.883 | 0.717 |
Honest tradeoff, stated plainly: v3's retention on 161_pass is the weakest of any AtmicEmbedding release — it has round-tripped to roughly base-model performance on this specific corpus (R@1 0.571 vs base's 0.592), losing most of the specialization v1→v2 training added there. This is a real cost of training on ~1.75x more data (3,176 vs the prior round's 1,808 rows) at the same epoch/learning-rate/freeze-layer budget, which pulls the model further from v2's weights. Notably, v2 itself underperforms base on the newer, more diverse own-test-data set (0.532 vs 0.611 accuracy) — v2 appears overfit to 161_pass's narrower style, and v3 trades that narrow strength for broader generalization. If 161_pass-specific retrieval quality matters for your use case, evaluate v3 there directly before switching from v2.