L6_bottom_distilled_v2
Lightweight multilingual sentence encoder optimized for intent classification.
Created from paraphrase-multilingual-MiniLM-L12-v2 via layer pruning + vocab pruning + improved knowledge distillation.
What's Different (v2)
This model uses an improved distillation strategy compared to v1:
| Hyperparameter | v1 | v2 |
|---|---|---|
| Learning rate | 2e-5 | 5e-6 (gentler) |
| Epochs | 3 | 10 (longer training) |
| MSE weight | 1.0 | 0.3 (less magnitude forcing) |
| Cosine weight | 0.5 | 2.0 (focus on direction alignment) |
Key insight: The original L6_bottom model already has strong syntactic representations from the teacher's bottom layers. Aggressive MSE loss destroys these. The v2 strategy uses cosine-dominant loss to align direction without forcing magnitude, preserving the model's existing strengths.
Results Comparison
| Model | MassiveIntent | MassiveScenario | Average |
|---|---|---|---|
| Teacher (12L, 480MB) | 55.52% | 61.01% | 58.27% |
| L6_bottom (no distill) | 54.70% | 59.39% | 57.05% |
| L6_bottom_distilled v1 | 51.63% | 59.03% | 55.33% |
| L6_bottom_distilled v2 | 53.06% | 59.86% | 56.46% |
v2 surpasses the original on MassiveScenarioClassification (59.86% > 59.39%).
Detailed Scores (v2)
| Language | MassiveIntent | MassiveScenario |
|---|---|---|
| ar | 42.95% | 48.90% |
| en | 61.44% | 69.26% |
| es | 56.33% | 62.55% |
| ko | 51.53% | 58.74% |
Model Details
| Property | Value |
|---|---|
| Teacher | paraphrase-multilingual-MiniLM-L12-v2 |
| Architecture | XLM-RoBERTa (pruned) |
| Hidden dim | 384 |
| Layers | 6 / 12 |
| Layer indices | [0, 1, 2, 3, 4, 5] |
| Strategy | Bottom half (syntactic-focused) |
| Vocab size | ~38,330 (pruned from 250K) |
| Safetensors | 98.1MB |
Quick Start
Training
Stage 1: Layer Pruning
- Teacher: paraphrase-multilingual-MiniLM-L12-v2 (12 layers)
- Selected layers: [0, 1, 2, 3, 4, 5] (bottom half, syntactic-focused)
- Vocab pruning: 250K -> ~38K tokens (corpus-based, 18 target languages)
Stage 2: Knowledge Distillation (v2)
- Loss: 0.3 * MSE + 2.0 * (1 - CosineSimilarity)
- Training data: MASSIVE dataset (90K multilingual sentences)
- Optimizer: AdamW (lr=5e-6, weight_decay=0.01)
- Schedule: Cosine annealing over 10 epochs
- Batch size: 64, GPU: NVIDIA RTX 5090
Limitations
- Vocabulary pruning restricts the model to 18 target languages
- Designed for short dialogue utterances, not long documents