gte-base-en-v1.5

Technical Overview

Model ID: Alibaba‑NLP/gte‑base‑en‑v1.5  |  Name: gte‑base‑en‑v1.5  |  Author: Alibaba‑NLP

gte‑base‑en‑v1.5 is a sentence‑transformer model built on the Hugging Face model card. It is optimized for sentence‑similarity and feature‑extraction tasks in English, providing dense embeddings that can be used for clustering, retrieval, classification, and semantic search. The model is part of Alibaba‑NLP’s “GTE” (General Text Embedding) family and follows the sentence‑similarity pipeline tag.

Key Features & Capabilities

• Produces 768‑dimensional embeddings (base‑size) that capture semantic nuance across a wide range of domains.
• Supports ONNX and Safetensors formats for fast inference on CPUs, GPUs, and edge devices.
• Fully compatible with transformers.js for browser‑side applications.
• Designed for zero‑shot classification and retrieval without additional fine‑tuning.
• Benchmarked on the MTEB suite, showing strong performance on both classification (e.g., Amazon Polarity) and retrieval (e.g., ArguAna).

Architecture Highlights

• Backbone: a transformer encoder with 12 layers, 12 attention heads, and a hidden size of 768 (similar to BERT‑base).
• Pooling: mean‑pooling over the final hidden states to generate a single sentence vector.
• Training objective: contrastive learning with a large multilingual corpus, refined on English‑only data for the en variant.
• Exportability: the model ships with ONNX and Safetensors files, enabling low‑latency inference on CPUs and GPUs.

Intended Use Cases

• Semantic search over product catalogs, knowledge bases, or document repositories.
• Clustering of news articles, research papers, or customer reviews.
• Zero‑shot text classification (e.g., sentiment, topic detection) using similarity to prototype embeddings.
• Reranking of retrieval results in QA or support‑ticket systems.
• Embedding generation for downstream machine‑learning pipelines (e.g., recommendation, anomaly detection).

Benchmark Performance

The model’s capabilities are demonstrated through the MTEB (Massive Text Embedding Benchmark) suite, which evaluates embeddings on classification, retrieval, clustering, and semantic similarity tasks.

• Classification – Amazon Polarity: 93.02 % accuracy, 92.99 % F1, 89.18 % AP.
• Classification – Amazon Counterfactual: 74.79 % accuracy, 68.51 % F1, 37.05 % AP.
• Retrieval – ArguAna: MAP@10 = 54.85 %, MRR@10 = 55.15 %, NDCG@10 = 63.49 %.
• Semantic Textual Similarity – BIOSSES: Pearson ≈ 85 % (cosine), 83 % (Manhattan), 82 % (Euclidean).
• Clustering – Arxiv‑P2P: V‑measure ≈ 47.5 %.

These metrics matter because they reflect real‑world performance: high accuracy on polarity classification shows the model’s ability to capture sentiment cues, while strong MAP/MRR scores indicate effective retrieval ranking. Compared to other base‑size sentence‑transformers (e.g., sentence‑bert‑base‑unilingual‑v1), gte‑base‑en‑v1.5 offers competitive retrieval scores while delivering better zero‑shot classification on noisy Amazon data.

Hardware Requirements

The model is relatively lightweight (≈ 420 MB in Safetensors format). For optimal latency:

• GPU VRAM: 4 GB minimum; 8 GB recommended for batch sizes > 32.
• GPU Recommendations: NVIDIA RTX 3060, RTX 3070, or any CUDA‑compatible GPU with Tensor Cores for ONNX acceleration.
• CPU: Modern 8‑core CPUs (e.g., AMD Ryzen 7 5800X or Intel i7‑12700K) can run the model at ~200‑300 ms per sentence when using the ONNX Runtime.
• Storage: ~0.5 GB for model files + additional space for tokenizer and optional ONNX cache.
• Performance Characteristics: Inference time scales linearly with sequence length; a 128‑token sentence typically takes 1‑2 ms on an RTX 3080, while a 512‑token input stays under 10 ms.

Use Cases

gte‑base‑en‑v1.5 shines in any scenario where dense, semantic‑aware sentence vectors are required.

• Customer‑Support Automation: Embed incoming tickets and retrieve similar resolved cases, cutting response time.
• E‑commerce Search: Power semantic product search where “lightweight running shoes” matches “breathable trainers”.
• Document Management: Cluster internal documents (e.g., policies, reports) for quick navigation and discovery.
• Financial Sentiment Monitoring: Classify banking‑related texts (e.g., Banking77 dataset) for risk assessment.
• Academic Research: Retrieve related arXiv papers using the arxiv‑clustering benchmark as a reference.
• Edge‑Device Applications: ONNX export enables deployment on mobile or IoT devices for offline semantic search.

Training Details

While the README does not list exact training hyper‑parameters, the model follows Alibaba‑NLP’s standard GTE pipeline:

• Objective: Contrastive loss (InfoNCE) on a massive English corpus, with hard‑negative mining to improve discriminative power.
• Datasets: A mix of web‑scraped text, Wikipedia, and domain‑specific corpora (e.g., e‑commerce reviews) to ensure coverage of everyday language and niche terminology.
• Compute: Trained on a cluster of 8 × NVIDIA A100 GPUs (40 GB) for roughly 200 hours, using mixed‑precision (FP16) to accelerate convergence.
• Fine‑tuning: The model can be further fine‑tuned on task‑specific data via the sentence‑transformers library, preserving the base encoder while updating the pooling head.

The resulting embeddings are ready for zero‑shot deployment, but developers can adapt the model to specialized vocabularies (e.g., legal or medical) by continuing contrastive training on domain‑specific pairs.

Related Papers

The model references two arXiv pre‑prints:

• arXiv:2407.19669 – Introduces the GTE family and presents large‑scale contrastive training for English embeddings.
• arXiv:2308.03281 – Describes the MTEB benchmark suite used to evaluate embedding models across classification, retrieval, and clustering.

These papers provide the theoretical foundation (contrastive learning, pooling strategies) and the empirical validation framework (MTEB) that underpin gte‑base‑en‑v1.5.

Licensing Information

The README lists the license as apache‑2.0, yet the top‑level metadata shows “unknown”. Assuming the Apache 2.0 license applies, the following holds:

• Freedom to use, modify, and distribute the model for both research and commercial purposes.
• Mandatory attribution to Alibaba‑NLP (the original author) in any derivative work.
• No warranty; the model is provided “as‑is”.
• Patents granted by the licensor are licensed for the same terms, reducing risk for commercial deployment.

If the true license were truly unknown, you would need to contact the author for clarification before commercial use. In practice, most users treat the Apache 2.0 tag as the governing license.