Time-Embed BGE-M3 LMS Temporal v1.9 Native-Day Bridge
This is a narrow corrective Time-Embed checkpoint for Korean LMS temporal retrieval.
It targets native Korean day-count offset failures, especially cases such as:
나흘 뒤vs4일 뒤닷새 전vs5일 전- adjacent hard negatives such as
나흘 뒤vs사흘 뒤
This checkpoint is not a broad data scaling run. It was trained as a minimal corrective bridge after v1.9 smoke evaluation found a specific remaining failure bucket.
Lineage
- Starting model:
kev-KOH/time-embed-bge-m3-lms-temporal-phase2-90 - Intermediate local model:
time-embed-bge-m3-native-day-focus-v1_9_4070ti/20260627-193859 - Uploaded checkpoint:
time-embed-bge-m3-native-day-bridge-v1_9_4070ti/20260627-195747 - Training rows:
data/bge_m3_temporal_native_day_bridge_v1_9/train.jsonl - Training size: 64 train rows, 32 dev rows, 32 test rows
Training Summary
- Model family: BGE-M3 / FlagEmbedding
- Pooling:
cls - Normalized embeddings: true
- Query max length: 128
- Passage max length: 128
- Train group size: 2
- Batch size: 16
- Gradient accumulation: 2
- Learning rate:
3e-7 - Temperature:
0.05 - Max steps: 10
- Hardware: RTX 4070 Ti SUPER 16GB
Evaluation
All comparisons below are same-runtime checks against kev-KOH/time-embed-bge-m3-lms-temporal-phase2-90 or the immediately preceding local focus checkpoint where noted.
v1.9 Smoke
| Model | margin_p10 | pairwise_accuracy | violation_rate | native_day_accuracy | native_day_violation | passed_gates |
|---|---|---|---|---|---|---|
| phase2-90 | -0.0722 | 0.8750 | 0.1250 | 0.5000 | 0.5000 | false |
| native-day focus step80 | -0.0011 | 0.8975 | 0.1025 | 0.5900 | 0.4100 | false |
| native-day bridge step10 | 0.0074 | 0.9600 | 0.0400 | 0.8400 | 0.1600 | true |
Broad Internal Checks
| Check | positive_p10 | critical_positive_score | margin_p10 | pairwise_accuracy | violation_rate | near_one_rate | passed_gates |
|---|---|---|---|---|---|---|---|
| semantic retention | 0.9793 | 0.9737 | 0.0523 | 1.0000 | 0.0000 | 0.1538 | true |
| temporal benchmark | 0.8505 | 0.7457 | 0.0408 | 1.0000 | 0.0000 | 0.0952 | true |
| calendar benchmark | 0.9241 | 0.8710 | 0.0316 | 0.9990 | 0.0010 | 0.0000 | true |
MTEB Forgetting Check
Same-runtime comparison against kev-KOH/time-embed-bge-m3-lms-temporal-phase2-90:
| Task | Metric | Base | Checkpoint | Delta |
|---|---|---|---|---|
| KLUE-STS | main_score | 0.884093 | 0.884490 | +0.000397 |
| KorSTS | main_score | 0.806877 | 0.807189 | +0.000312 |
| KLUE-NLI | main_score | 0.725636 | 0.723709 | -0.001927 |
| SQuADKorV1Retrieval | ndcg_at_10 | 0.899780 | 0.900000 | +0.000220 |
| SQuADKorV1Retrieval | mrr_at_10 | 0.875715 | 0.875811 | +0.000096 |
The largest negative MTEB movement is KLUE-NLI.main_score = -0.001927, which is inside the default retention band used for Time-Embed promotion checks.
Intended Use
Use this model for Korean LMS retrieval experiments where temporal expressions are important, especially course materials, assignment deadlines, notices, and lecture timing.
The model is optimized for temporal distinction, not as a general-purpose Korean embedding replacement. For broad tasks, compare against BAAI/bge-m3 and the Time-Embed baseline before deployment.
Limitations
- This is a narrow corrective checkpoint, not a full temporal ontology expansion.
- Native-day offset coverage is improved but should still be stress-tested for
사흘,나흘,닷새,엿새,이레,여드레,아흐레, and열흘. - Calendar benchmark margin slack decreased relative to the phase2-90 baseline, although gates passed.
- Semantic retention
near_one_rateincreased while remaining below the configured gate; future runs should monitor collapse risk.