kev-KOH/time-embed-bge-m3-lms-temporal-v1-6-1

Time Embed BGE-M3 LMS Temporal v1.6.1

Fine-tuned BGE-M3 checkpoint for Korean LMS temporal query embedding.

This model targets equivalence and contrast around Korean relative-time expressions, for example 일주일 전, 7일 전, and 1주 전 as positives, while separating hard negatives such as 지난주 or 최근 7일 when the temporal meaning differs.

Source

• Base model: BAAI/bge-m3
• Training run: runs/time-embed-bge-m3-v1_6_1/20260524-162559
• Training data: kev-KOH/time-embed-bge-m3, folder bge_m3_temporal_focus_v1_6_1
• Fine-tuning setup: batch 32, train group size 8, bf16, CLS pooling, normalized embeddings

Evaluation

Evaluated with the project v1.6.1 temporal gates on 3,000-row dev/test subsets.

split	pairwise_accuracy	hard_negative_violation_rate	margin_mean	margin_p10	random_pair_p95	near_one_rate	passed_gates
dev	0.981000	0.019000	0.181198	0.060985	0.497738	0.001343	True
test	0.980667	0.019333	0.179639	0.059535	0.522146	0.001307	True

Full eval JSON files are included under eval_reports/.

Intended Use

Use for Korean LMS retrieval/reranking pipelines where temporal semantics must distinguish exact offsets, deictic expressions, duration expressions, windows, and deadline boundaries.

Limitations

This is a focused temporal adaptation. Validate on downstream LMS retrieval traffic before production promotion.