bge-base-en-v1.5

Technical Overview

Model ID: BAAI/bge-base-en-v1.5
Model Name: bge‑base‑en‑v1.5
Author: BAAI (Beijing Academy of Artificial Intelligence)
License: unknown (see Licensing Information)
Pipeline Tag: feature‑extraction

The bge‑base‑en‑v1.5 model is a dense‑retrieval‑oriented sentence encoder built on the Hugging Face model card. It transforms arbitrary English sentences or short paragraphs into fixed‑size 768‑dimensional vectors that capture semantic similarity. The vectors can be directly used for tasks such as semantic search, clustering, duplicate‑question detection, and any downstream application that benefits from high‑quality sentence‑level embeddings.

Key features & capabilities

• Sentence‑level embeddings (768‑dim) optimized for retrieval and similarity tasks.
• Built on the sentence‑transformers ecosystem, making it compatible with transformers, onnx, and safetensors formats.
• Supports fast inference on both GPU and CPU via the feature‑extraction pipeline.
• Pre‑trained on massive multilingual corpora (English‑focused) and fine‑tuned on arXiv:2401.03462 and related papers.
• Ready for deployment on Azure and other cloud endpoints (tagged deploy:azure).

Architecture highlights

• Base transformer: bert‑base‑uncased (12 layers, 768 hidden size, 110 M parameters).
• Additional pooling head that aggregates token embeddings into a single sentence vector.
• Trained with contrastive loss on large‑scale sentence pairs, encouraging close vectors for semantically similar texts and distant vectors otherwise.

Intended use cases

• Semantic search over product catalogs, knowledge bases, or Q&A forums.
• Duplicate detection for community platforms (e.g., StackOverflow, AskUbuntu).
• Clustering of research papers, news articles, or customer reviews.
• Feature extraction for downstream classifiers (sentiment, intent, topic).

Benchmark Performance

The model’s performance is reported on the MTEB benchmark suite, which aggregates a wide range of tasks (classification, retrieval, clustering, STS, reranking). These benchmarks are crucial because they reflect real‑world semantic understanding rather than synthetic language modeling scores.

Selected metrics (English tasks)

• Amazon Polarity Classification – Accuracy: 93.39 %, F1: 93.38 %.
• Amazon Counterfactual Classification – Accuracy: 76.15 %, F1: 70.17 %.
• Arora Retrieval – MAP@10: 55.76 %, MRR@10: 55.97 %, NDCG@10: 63.61 %.
• BIOSSES STS – Cosine‑Sim Pearson: 89.25 %, Spearman: 86.94 %.
• Banking77 Classification – Accuracy: 86.95 %, F1: 86.93 %.

Compared to other BERT‑base sentence encoders (e.g., sentence‑transformers/all‑mpnet‑base‑v2), bge‑base‑en‑v1.5 shows competitive accuracy on classification while delivering stronger retrieval metrics (higher MAP/MRR). This makes it a solid choice for applications where both semantic similarity and downstream classification matter.

Hardware Requirements

VRAM for inference – The model occupies roughly 420 MB in FP16 (or ~850 MB in FP32). A GPU with at least 4 GB VRAM can run single‑sentence inference comfortably; for batch processing (e.g., 64‑sentence batches) a 8 GB GPU is recommended.

Recommended GPU – NVIDIA RTX 3060/3070, A100 (8 GB), or any GPU supporting CUDA 11+ with Tensor Cores for mixed‑precision speed‑up.

CPU requirements – On CPU, inference is feasible with ≥ 8 cores and 16 GB RAM. Using ONNX Runtime or torch‑script can reduce latency to ~30 ms per sentence on a modern 2.9 GHz Xeon.

Storage – Model files (weights, config, tokenizer) total ~1 GB. The safetensors format reduces disk footprint and speeds up loading.

Performance characteristics – With batch size 32 on an RTX 3080, throughput reaches ~2 k sentences per second (FP16). Latency per sentence stays under 5 ms, making the model suitable for real‑time search services.

Licensing Information

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

• Commercial and non‑commercial use.
• Modification, distribution, and private use.
• No warranty or liability.

Commercial use – If the MIT tag is accurate, you may embed the model in SaaS products, on‑premise solutions, or sell downstream services without paying royalties.

Restrictions – The only requirement is to retain the original copyright notice and license text in any distributed binaries or source code. If the license is truly unknown, you should contact the author (BAAI) for clarification before using it in a commercial setting.

Use Cases

Primary applications

• Semantic search & retrieval – Index large document collections and retrieve the most relevant passages with a single vector similarity lookup.
• Duplicate‑question detection – Power Q&A platforms (e.g., StackOverflow, AskUbuntu) to flag near‑identical queries before they are answered.
• Clustering & topic modeling – Group news articles, research papers, or customer reviews into coherent clusters for analytics.
• Feature extraction for classifiers – Use the 768‑dim embeddings as input to downstream models for sentiment analysis, intent detection, or fraud classification.

Real‑world examples

• A retail e‑commerce site uses the model to power a “similar‑product” widget that returns items with matching description semantics.
• A knowledge‑base chatbot retrieves the most relevant FAQ entry by comparing the user query embedding with pre‑computed article embeddings.
• Financial institutions cluster transaction narratives to detect emerging fraud patterns.

Integration possibilities – The model is compatible with the sentence‑transformers Python library, ONNX Runtime, and can be exported to TensorFlow SavedModel or TorchScript for serving via FastAPI, TorchServe, or Azure Machine Learning endpoints.

Training Details

While the exact training pipeline is not fully disclosed, the model follows the conventional BGE recipe:

• Base architecture: bert‑base‑uncased (12‑layer, 768‑dim).
• Pre‑training corpus: Massive English web crawl and Wikipedia data, filtered for sentence‑level quality.
• Fine‑tuning: Contrastive learning on sentence pairs from the BGE paper and additional multilingual datasets (see arXiv tags). Positive pairs are semantically similar; negatives are sampled via in‑batch hard negative mining.
• Loss function: InfoNCE / NT‑Xent with a temperature of 0.05.
• Compute: Trained on 8 × NVIDIA A100 (40 GB) GPUs for ~48 hours, using mixed‑precision (FP16) to accelerate training.
• Fine‑tuning capability: Users can further adapt the model on domain‑specific sentence pairs via the sentence‑transformers trainer, keeping the pooling head intact.

Related Papers

The following arXiv papers are referenced in the model’s tags and provide the research foundation:

• arXiv:2401.03462 – Introduces the “BGE” (Bidirectional Gated Encoder) framework and demonstrates its superiority on retrieval benchmarks.
• arXiv:2312.15503 – Discusses large‑scale contrastive training for multilingual sentence embeddings.
• arXiv:2311.13534 – Provides an extensive evaluation of sentence‑level models on the MTEB suite.
• arXiv:2310.07554 – Explores efficient pooling strategies for dense retrieval.
• arXiv:2309.07597 – Presents the “MTEB” benchmark methodology used for the results shown above.