Time-Embed BGE-M3 LMS Temporal Phase2.72
This model is a Korean temporal-expression embedding checkpoint for LMS retrieval. It starts from the Time-Embed phase2.67 checkpoint and applies a small multi-objective pairwise update with train-only strict-frontier analogs.
HF repo: kev-KOH/time-embed-bge-m3-lms-temporal-phase2-72
Intended Use
Use this checkpoint for Korean LMS retrieval queries where relative temporal expressions must be separated precisely, for example:
일주일 전,7일 전,1주 전as exact duration offsets지난주as a calendar week period최근 7일as a rolling window다음 주versus7일 뒤
The model is not a general-purpose replacement for BAAI/bge-m3; it is a
specialized temporal retrieval checkpoint. Run downstream retrieval validation
before production use.
Training Method
- Base/source checkpoint:
research/runs/time_embed_phase2_67_p65_uploaded_material_calendar_tail_48_4070ti/20260626-071336 - Objective: multi-objective pairwise margin with positive-floor, negative-ceiling, teacher score-MSE, and PCGrad
- Trainable layers: last two encoder layers
- Learning rate:
5e-9 - Steps:
80 - Precision: fp32
- Training shards:
- temporal exact-week frontier analogs:
50rows - semantic year-offset frontier analogs:
48rows - uploaded-material calendar frontier analogs:
72rows
- temporal exact-week frontier analogs:
Internal Strict-Frontier Evaluation
| suite | strict frontier | phase2.72 critical positive | delta | passed gates | negative near-one |
|---|---|---|---|---|---|
| temporal benchmark | 0.714914416 | 0.715032358 | +0.000117942 | true | 0.000000 |
| semantic retention | 0.961084034 | 0.961094075 | +0.000010041 | true | 0.000000 |
| calendar benchmark | 0.874667996 | 0.874748229 | +0.000080233 | true | 0.000000 |
Caveat: the semantic strict-frontier gain is very small (+0.000010041). Treat
this checkpoint as a narrow frontier improvement and continue to monitor semantic
retention in downstream data.
MTEB Catastrophic Forgetting Check
Compared against BAAI/bge-m3 in the same runtime:
| task | metric | base | checkpoint | delta |
|---|---|---|---|---|
| KLUE-STS | main_score | 0.877121 | 0.884086 | +0.006965 |
| KorSTS | main_score | 0.802612 | 0.806875 | +0.004263 |
| KLUE-NLI | main_score | 0.700598 | 0.725633 | +0.025035 |
| SQuADKorV1Retrieval | ndcg_at_10 | 0.904110 | 0.899780 | -0.004330 |
| SQuADKorV1Retrieval | mrr_at_10 | 0.880192 | 0.875715 | -0.004477 |
No general catastrophic forgetting was detected by the project default rule: all
drops stayed within the -0.01 retention band.
Usage
from FlagEmbedding import BGEM3FlagModel
model = BGEM3FlagModel(
"kev-KOH/time-embed-bge-m3-lms-temporal-phase2-72",
use_fp16=True,
pooling_method="cls",
)
embeddings = model.encode(
["7일 전 업로드된 데이터 자료 찾아줘", "지난주 업로드된 데이터 자료 찾아줘"],
return_dense=True,
return_sparse=False,
return_colbert_vecs=False,
)["dense_vecs"]
Artifacts
model.safetensorsconfig.jsontokenizer.jsontokenizer_config.jsonmulti_objective_pairwise_training_summary.json