contrastive objective concept
| Definition | A training rule that pulls pairs declared similar together and pushes other pairs apart. The declared pairs are where an encoder learns what similar means. Chapter 12. Also contrastive, declared similar, fine-tun. |
|---|---|
| Example | Two paraphrases are pulled to cosine near one and an unrelated pair pushed toward zero, and the choice of pairs decides what the encoder can tell apart. |
| Book | Data Mining as Observation, draft 0.2, commit f3914f0; entry id contrastive-objective, kind concept. |
| Status | no ledger row names this entry. Corrections: none recorded. |
| Defining equation | none |
| Assumptions and scope |
|
| Prior art | none recorded |
| Evidence | lean/DataMiningAsObservation/Contrastive.lean |
| Reviewed | not yet reviewed; generated 2026-09-10 from records at the commits on the provenance page. |
Equation
none
Conditions
- A training rule that pulls pairs declared similar together and pushes other pairs apart. With a margin loss the loss is nonnegative, zero exactly when every declared pair scores above every other pair by the margin, and never lower for a larger margin.
- Swapping which pairs are declared similar reverses that condition, and with a positive margin no scorer satisfies both, so the declared pairs are where an encoder learns what similar means. Fine-tuning moves the reader, and the cross-corpus positives were the fix when within-corpus scores collapsed.
Conditions are curated in entries.toml rather than read
from a record.
Ledger
none
First stated
Chapter 12 section 12.4 of Data Mining as Observation, with
the program’s encoders in xbse, xbse/README.md:104-130.
Measurements
none
Failures and corrections
none
Invariance envelope
none declared
Machine checked
lean/DataMiningAsObservation/Contrastive.lean,
theorems loss_nonneg, loss_eq_zero_iff,
not_both, loss_mono_margin, at
observation-data-mining f3914f0; what the check covers is stated in the
book’s appendix
C.
Used in
Data Mining as Observation chapters 0, 11, 12.
Related
encoder; embedding; cross-corpus gate; quotient.
See also
Book equations stated beside the entry’s terms, not defining it: 12.3, 11.1.
Ledger rows that cite the entry’s records without naming it: NEG-2, GO-B-legal (035→036).
Sources-table rows that share a record with the entry without naming it: chapter 12 section 12.5.
Status
Generated 2026-09-10 by encyclopedia/generate.py; book
at observation-data-mining f3914f0; the commit of every record is listed
in the encyclopedia’s provenance.