kev-KOH/time-embed-bge-m3-lms-temporal-phase2-41

Time-Embed BGE-M3 LMS Temporal Phase2.41

This model is a Time-Embed research checkpoint for Korean LMS temporal-expression retrieval. It is based on BAAI/bge-m3 and was selected after Phase2.41 corrected the collapse gate to distinguish true positive temporal alignment from negative or global embedding saturation.

Intended Use

Use this checkpoint for Korean LMS queries where temporally equivalent expressions should be close in embedding space, for example:

• 일주일 전, 7일 전, 1주 전
• 이번 주, 이번주, 금주
• 지난 7일, 최근 7일

The model is intended for retrieval/ranking experiments, not as a general-purpose replacement for every Korean embedding workload.

Source Checkpoint

Local training artifact:

research/runs/time_embed_phase2_40_boundary_context_target_compression_160_4070ti/20260626-005203

Phase2.41 did not train a new checkpoint. It reclassified the Phase2.40 checkpoint after row-level diagnostics showed that the old aggregate near_one_rate gate was penalizing positive-only temporal alignment.

Internal Evaluation

suite	critical positive	negative p95	margin p10	negative near-one	random-pair p95	passed
temporal benchmark, corrected gate	0.712765	0.927091	0.061491	0.000000	0.525630	true
focused assignment/quiz	0.917938	0.935560	0.044867	0.000000	0.618668	true
calendar benchmark	0.873677	0.912210	0.050455	0.000000	0.624696	true
semantic retention	0.960865	0.934735	0.056170	0.000000	0.768753	true

Important diagnostic:

• aggregate temporal near_one_rate: 0.067178
• positive temporal near_one_rate: 0.135700
• negative temporal near_one_rate: 0.000000

The high near-one mass is positive-only and corresponds to true temporal equivalence surfaces, so it is not treated as collapse.

MTEB Forgetting Check

Base model: BAAI/bge-m3

task	metric	base	checkpoint	delta
KLUE-STS	main_score	0.877152	0.884022	+0.006870
KorSTS	main_score	0.802649	0.806705	+0.004056
KLUE-NLI	main_score	0.700609	0.725655	+0.025046
SQuADKorV1Retrieval	ndcg_at_10	0.904150	0.899820	-0.004330
SQuADKorV1Retrieval	mrr_at_10	0.880194	0.875764	-0.004430

No general catastrophic forgetting was detected under the project thresholds.

Loading

from FlagEmbedding import BGEM3FlagModel

model = BGEM3FlagModel(
    "kev-KOH/time-embed-bge-m3-lms-temporal-phase2-41",
    use_fp16=True,
)
embeddings = model.encode(
    ["일주일 전 업로드된 데이터 자료 찾아줘", "7일 전 업로드된 데이터 자료 찾아줘"],
    return_dense=True,
    return_sparse=False,
    return_colbert_vecs=False,
)["dense_vecs"]

Provenance

Project repository: Han-taz/Time-Embed

Relevant local documentation:

• wiki/2026-06-26/research/phase2-41-collapse-gate-correction-design.md
• wiki/2026-06-26/eval-results/phase2-41-collapse-gate-corrected-phase2-40/metrics-summary.md
• wiki/2026-06-26/eval-results/phase2-41-collapse-gate-corrected-phase2-40/mteb-forgetting/README.md