Property	Value
Teacher	paraphrase-multilingual-MiniLM-L12-v2
Architecture	XLM-RoBERTa (pruned)
Hidden dim	384
Layers	4 (from 12)
Layer indices	[8, 9, 10, 11]
Strategy	4 layers, top quarter (semantic-focused compact)
Est. params	103,283,328
Est. FP32	394.0MB
Est. INT8	98.5MB
Est. INT8 + vocab pruned	27.1MB

Supported Languages (18)

ko, en, ja, zh, es, fr, de, pt, it, ru, ar, hi, th, vi, id, tr, nl, pl

Intended Use

This is a student encoder designed to be used as the backbone for a lightweight 3-class intent classifier (Action / Recall / Other) in multilingual dialogue systems.

Action: User requests an action (book, order, change settings, etc.)
Recall: User asks about past events or stored information
Other: Greetings, chitchat, emotions, etc.

Usage

python

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("L4_top")
embeddings = model.encode(["예약 좀 해줘", "지난번 주문 뭐였지?", "안녕하세요"])
print(embeddings.shape)  # (3, 384)

MTEB Results

MassiveIntentClassification

Average: 45.26%

Language	Score
ar	34.77%
en	55.68%
es	42.64%
ko	47.96%

MassiveScenarioClassification

Average: 48.79%

Language	Score
ar	36.89%
en	61.66%
es	47.3%
ko	49.3%

Training / Distillation

This model was created via layer pruning (no additional training):

Load teacher: paraphrase-multilingual-MiniLM-L12-v2 (12 layers, 384 hidden)
Select layers: [8, 9, 10, 11]
Copy embedding weights + selected layer weights
Wrap with mean pooling for sentence embeddings

For deployment, vocabulary pruning (250K → ~55K tokens) and INT8 quantization are applied to meet the ≤50MB size constraint.

Limitations

Layer pruning without fine-tuning may lose some quality vs. proper knowledge distillation
Vocabulary pruning limits the model to the target 18 languages
Designed for short dialogue utterances, not long documents

gomyk/intent-student-L4_top

Intent Classifier Student: L4_top

Model Details