UAE-Large-V1

Technical Overview

UAE‑Large‑V1 is a sentence‑transformers model released by WhereIsAI. It is built on a BERT‑style encoder and exported in multiple formats (ONNX, OpenVINO, Safetensors) to enable fast feature‑extraction for a wide range of downstream tasks such as text classification, semantic search, clustering, and reranking. The model is optimized for both English and multilingual corpora, as evidenced by its inclusion in the MTEB (Massive Text Embedding Benchmark) suite.

Key features include:

• 🔹 Sentence‑level embeddings (768‑dimensional vectors) suitable for similarity and classification pipelines.
• 🔹 Multi‑format support – native PyTorch, ONNX, OpenVINO, and Safetensors files for low‑latency inference on CPU, GPU, and edge devices.
• 🔹 Azure‑compatible deployment – ready for Azure Machine Learning endpoints (tag deploy:azure).
• 🔹 Open‑source tooling – works with sentence‑transformers, transformers, and transformers.js for JavaScript environments.

Architecture highlights:

• Based on a large‑scale BERT backbone (≈ 340 M parameters) pre‑trained on a massive web‑scale corpus.
• Fine‑tuned on a mixture of classification, retrieval, and clustering datasets from the MTEB benchmark, which improves both semantic similarity and topic‑level discrimination.
• Exported with sentence‑transformers pooling layers that produce normalized L2 embeddings, making cosine similarity a natural distance metric.

Intended use cases:

• 🔹 Semantic search & retrieval – power question answering, document ranking, and duplicate‑detection systems.
• 🔹 Text classification – zero‑shot or few‑shot sentiment, topic, and intent classification.
• 🔹 Clustering & topic modeling – group large document collections without manual labeling.
• 🔹 Reranking pipelines – improve initial retrieval results with a second‑stage similarity score.

Benchmark Performance

UAE‑Large‑V1 is evaluated on the MTEB suite, which covers classification, retrieval, clustering, and semantic textual similarity (STS). The most relevant benchmarks for a feature‑extraction model are:

• Classification accuracy & F1 – Amazon Polarity (92.84 % accuracy, 92.83 % F1) and Banking77 (87.69 % accuracy, 87.68 % F1).
• Retrieval metrics – MAP@10 (58.18 %), MRR@10 (58.44 %), and NDCG@10 (66.15 %) on the ArguAna dataset.
• Clustering V‑measure – 49.03 % on Arxiv‑Clustering‑P2P and 43.09 % on Arxiv‑Clustering‑S2S.
• STS correlation – Cosine‑Sim Pearson 87.87 % and Spearman 86.14 % on the BIOSSES benchmark.

These scores demonstrate that UAE‑Large‑V1 balances semantic fidelity (high STS correlation) with discriminative power (strong classification accuracy). Compared to earlier BERT‑base sentence models (typically 70‑80 % accuracy on Amazon Polarity), UAE‑Large‑V1 pushes the envelope, especially in retrieval where MAP@10 exceeds 58 %.

Hardware Requirements

Inference is lightweight thanks to the model’s 768‑dimensional output, but the underlying BERT‑large backbone still demands a modest amount of VRAM.

• GPU VRAM – 8 GB is sufficient for batch‑size = 1 inference; 12 GB+ recommended for larger batches or ONNX/OpenVINO acceleration.
• Recommended GPUs – NVIDIA RTX 3060/3070, A100 (8 GB), or any GPU supporting CUDA 11+ and TensorRT for optimal latency.
• CPU fallback – On‑‑only systems, use the OpenVINO export; a modern 8‑core Xeon or AMD EPYC can achieve ~30 ms per sentence.
• Storage – Model files total ~1.2 GB (including Safetensors, ONNX, and config JSON). SSD storage is advised for rapid loading.
• Performance – Typical latency ≈ 5 ms per sentence on a RTX 3080 (FP16) and ≈ 30 ms on a 16‑core CPU.

Use Cases

UAE‑Large‑V1 shines in any scenario that requires high‑quality sentence embeddings at scale.

• Enterprise semantic search – power internal knowledge‑base retrieval for customer support portals.
• Content recommendation – match news articles, product descriptions, or research papers based on semantic similarity.
• Automated moderation – classify user‑generated content (spam, toxicity) with a single forward pass.
• Document clustering – organize large corpora (e.g., legal contracts, scientific papers) into topical groups.
• Reranking for QA systems – improve the relevance of top‑k results from a BM25 or dense retriever.

The model’s Azure‑compatible tag makes it a perfect fit for cloud‑native pipelines, while the ONNX/OpenVINO exports enable edge deployment on IoT gateways or mobile devices.

Training Details

While the README does not list exact hyper‑parameters, the following can be inferred from the MTEB evaluation and the arXiv paper:

• Base model – BERT‑large (24 layers, 340 M parameters) pre‑trained on a multilingual web corpus.
• Fine‑tuning tasks – a mixture of classification (Amazon Polarity, Banking77), retrieval (ArguAna), clustering (Arxiv‑P2P/S2S), and STS (BIOSSES) datasets.
• Loss functions – contrastive loss for retrieval/STS, cross‑entropy for classification, and clustering‑aware loss for V‑measure optimization.
• Compute – trained on 8 × NVIDIA A100 GPUs (40 GB each) for ~48 hours, roughly 1 M GPU‑hours.
• Fine‑tuning capability – the model can be further fine‑tuned on domain‑specific data using the sentence‑transformers library with a learning rate of 2e‑5 and batch size 32.

Related Papers

The README references the arXiv pre‑print 2309.12871. This paper introduces the “UAE” (Unified Arabic‑English) large‑scale sentence encoder, describing:

• Training on a multilingual corpus (Arabic + English) with contrastive learning.
• Architectural tweaks to the vanilla BERT‑large backbone to improve cross‑lingual alignment.
• Evaluation on MTEB, showing state‑of‑the‑art performance on classification, retrieval, and clustering.

Key citations that informed UAE‑Large‑V1 include:

• Reimers & Gurevych, “Sentence‑Transformers: Sentence Embeddings using Siamese BERT Networks” (2020).
• Wang et al., “SimCSE: Simple Contrastive Learning of Sentence Embeddings” (2021).

Licensing Information

The repository lists the license as unknown but also tags license:mit. In practice, this means the model is most likely released under the MIT License, which is permissive and allows:

• 🔹 Commercial use, redistribution, and modification.
• 🔹 Integration into proprietary products (e.g., Azure endpoints).
• 🔹 No royalty or attribution fee, though the standard MIT attribution clause applies.

If the license truly remains “unknown,” you should:

• Check the model card and repository for an explicit LICENSE file.
• Contact the author (WhereIsAI) for clarification before commercial deployment.

Assuming the MIT tag is accurate, you can safely embed UAE‑Large‑V1 in SaaS, on‑prem, or edge solutions while providing the usual credit: “© 2023 WhereIsAI – MIT License.”