SBERT-all-MiniLM-L6-with-pooler
| Property | Value |
|---|---|
| Author | optimum |
| Vector Dimension | 384 |
| Base Model | MiniLM-L6-H384-uncased |
| Training Data | 1B+ sentence pairs |
| Max Sequence Length | 256 tokens |
What is sbert-all-MiniLM-L6-with-pooler?
SBERT-all-MiniLM-L6-with-pooler is an ONNX-optimized sentence embedding model that maps text to dense 384-dimensional vectors. Unlike standard SBERT models, this version includes both last_hidden_state and pooler_output, making it more versatile for downstream tasks. The model was trained on over 1 billion sentence pairs using contrastive learning, making it particularly effective for semantic search, clustering, and similarity tasks.
Implementation Details
The model is built on MiniLM-L6-H384-uncased and fine-tuned using a contrastive learning objective. Training utilized TPU v3-8 hardware with a batch size of 1024 over 100k steps. The implementation uses the AdamW optimizer with a 2e-5 learning rate and includes a 500-step warmup period.
- ONNX-optimized for efficient inference
- Dual output architecture (hidden states and pooler)
- Trained on diverse datasets including Reddit, WikiAnswers, and academic citations
- Optimized for sentences and short paragraphs up to 256 tokens
Core Capabilities
- Semantic text embedding generation
- Information retrieval and clustering
- Sentence similarity computation
- Cross-lingual text comparison
- Document classification support
Frequently Asked Questions
Q: What makes this model unique?
This model stands out due to its ONNX optimization and dual output architecture, providing both hidden states and pooler outputs. It's trained on an exceptionally large and diverse dataset of over 1 billion sentence pairs, making it robust for various text similarity tasks.
Q: What are the recommended use cases?
The model excels in semantic search applications, document clustering, similarity analysis, and information retrieval tasks. It's particularly effective for short to medium-length text processing, with optimal performance on content under 256 tokens.
