Time-Embed BGE-M3 LMS Temporal Phase2.74
This checkpoint is a Korean LMS temporal-expression embedding model derived
from BAAI/bge-m3.
Phase2.74 is not a new gradient-training run. It is a training-free convex interpolation between two Time-Embed checkpoints:
candidate = (1 - alpha) * phase2.72 + alpha * phase2.73
alpha = 0.75
The purpose is to keep the Phase2.72 temporal frontier while borrowing the small semantic/calendar gains observed in Phase2.73.
Intended Use
Use this model for Korean LMS retrieval where relative temporal expressions must be separated precisely, for example:
일주일 전,7일 전,1주 전지난주versus7일 전최근 7일versus7일 전지난달versus30일 전- same time expression but different LMS item type or intent
This model is specialized for temporal-expression retrieval. It is not a
general-purpose replacement for BAAI/bge-m3 without task-specific validation.
Internal Evaluation
Promotion baseline: Phase2.72.
| suite | Phase2.72 frontier | Phase2.74 | delta |
|---|---|---|---|
| temporal benchmark critical positive | 0.715032358 | 0.715032517 | +0.000000159 |
| semantic retention critical positive | 0.961094075 | 0.961106231 | +0.000012156 |
| calendar benchmark critical positive | 0.874748229 | 0.874759797 | +0.000011568 |
All three internal suites passed their gates, and all three suites had
negative_near_one_rate=0.0.
MTEB Retention Check
Base-vs-checkpoint Korean MTEB comparison:
| task | metric | delta |
|---|---|---|
| KLUE-STS | main_score | +0.006886 |
| KorSTS | main_score | +0.004257 |
| KLUE-NLI | main_score | +0.024984 |
| SQuADKorV1Retrieval | ndcg_at_10 | -0.004290 |
| SQuADKorV1Retrieval | mrr_at_10 | -0.004378 |
No general catastrophic forgetting was detected by the project default rule:
all drops stayed within the -0.01 retention band.
Caveat
The smallest Phase2.74 frontier gain over Phase2.72 is only +0.000000159.
Treat this model as a narrow frontier improvement, not a wide-margin quality
jump. Downstream consumers should keep Phase2.72 as a fallback candidate until
Phase2.74 is validated in the target LMS retrieval stack.
Provenance
- Base model:
BAAI/bge-m3 - Parent checkpoints:
kev-KOH/time-embed-bge-m3-lms-temporal-phase2-72- local Phase2.73 checkpoint
- Research note:
wiki/2026-06-26/research/phase2-74-p72-p73-semantic-frontier-interpolation-design.md