Phase2.90 is a tiny p74-parent continuation run. It keeps the uploaded Phase2.74 checkpoint as both base and teacher, then applies a 14-step multi-objective pairwise update with stronger semantic proxy preservation.

The goal is narrow: preserve Phase2.74 semantic/calendar frontiers while adding a small temporal lift.

Intended Use

Use this model for Korean LMS retrieval where relative temporal expressions must be separated precisely, for example:

일주일 전, 7일 전, 1주 전
지난주 versus 7일 전
최근 7일 versus 7일 전
지난달 versus 30일 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.

Training Summary

item	value
base/teacher	`phase2.74`
learning rate	`1.5e-10`
max steps	`14`
trainable scope	`encoder.layer.23.*`
precision	`fp32`
gradient surgery	`pcgrad`
seed	`90`

Train-only shards:

shard	rows
calendar proxy guard	`48`
temporal hard repair	`144`
positive anchor replay	`120`
temporal concept replay	`80`
semantic year-offset guard	`24`
visible semantic control replay	`16`

Internal Evaluation

Promotion baseline: Phase2.74.

suite	Phase2.74 frontier	Phase2.90	delta
temporal benchmark critical positive	`0.715032517`	`0.715032874`	`+0.000000358`
semantic retention critical positive	`0.961106231`	`0.961106708`	`+0.000000477`
calendar benchmark critical positive	`0.874759797`	`0.874762101`	`+0.000002305`

All three final-suite gates passed and negative_near_one_rate=0.0 in the temporal, semantic, and calendar suites.

Paired p74 audit:

suite	verdict	critical delta
temporal benchmark	`strict_pass_but_fragile_tie`	`+0.000000358`
semantic retention	`strict_pass_but_fragile_tie`	`+0.000000477`
calendar benchmark	`strict_pass_but_fragile_tie`	`+0.000002305`

MTEB Retention Check

Base-vs-checkpoint Korean MTEB comparison:

task	metric	delta
KLUE-STS	main_score	`+0.006886`
KorSTS	main_score	`+0.004260`
KLUE-NLI	main_score	`+0.024989`
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

Phase2.90 is a narrow frontier improvement over Phase2.74, not a wide-margin quality jump. Treat this model as the current best candidate only with the project's internal temporal/calendar/semantic gates attached.

Provenance

Base model: BAAI/bge-m3
Parent checkpoint: kev-KOH/time-embed-bge-m3-lms-temporal-phase2-74
Research note: wiki/2026-06-26/research/phase2-90-p74-semantic-proxy-guarded-lift-design.md

kev-KOH/time-embed-bge-m3-lms-temporal-phase2-90

Time-Embed BGE-M3 LMS Temporal Phase2.90