snowflake-arctic-embed2

Snowflake 的前沿嵌入模型。Arctic Embed 2.0 增加了多語言支援，同時不犧牲英文效能或可擴展性。

嵌入 568m

3.72萬下載次數已更新 3個月前

3 個標籤

3個月前更新

3個月前

5de93a84837d · 1.2GB

Apache License Version 2.0, January 200

11kB

說明文件

雪花很高興宣布 Arctic Embed 2.0 的發布，這是我們前沿嵌入模型的下一次迭代，現在可以支援多語言搜尋。雖然我們之前的版本受到了客戶、合作夥伴和開源社群的好評，並帶來了數百萬次的下載，但我們一直收到一個要求：你們可以讓這個模型支援多語言嗎？ Arctic Embed 2.0 建立在我們先前版本的穩固基礎之上，增加了多語言支援，同時不犧牲英文效能或可擴展性，以滿足更廣泛的使用者群體的需求，這些使用者群體涵蓋了廣泛的語言和應用程式。

圖 1. 參數少於 10 億的開源多語言嵌入模型的單向量密集檢索效能。分數是 MTEB 檢索和 CLEF (ELRA, 2006) 子集（涵蓋英文、法文、西班牙文、義大利文和德文）的平均 nDCG@10。

Arctic Embed 2.0 多樣且強大的功能集

企業級的吞吐量和效率：Arctic Embed 2.0 模型是為大規模企業需求而建構的。即使是我們的「大型」模型，參數也遠低於 10 億，並提供快速、高吞吐量的嵌入能力。根據內部測試，它在 NVIDIA A10 GPU 上輕鬆處理每秒超過 100 個文檔（平均），並實現低於 10 毫秒的查詢嵌入延遲，從而在經濟實惠的硬體上實現實際部署。
英文和非英文檢索的卓越品質：儘管尺寸緊湊，但 Arctic Embed 2.0 模型在各種英文和非英文基準資料集上都取得了令人印象深刻的 NDCG@10 分數，展現了即使對於未包含在訓練配方中的語言也能很好地泛化的能力。這些令人印象深刻的基準分數使 Arctic Embed 2.0 成為前沿檢索模型中的領導者。
透過 Matryoshka Representation Learning (MRL) 實現可擴展的檢索：Arctic Embed 2.0 版本包含與 Arctic Embed 1.5 中引入的相同的量化友善 MRL 功能，允許使用者在對大型資料集執行搜尋時降低成本並優化規模。透過兩種模型尺寸，使用者可以使用每個向量僅 128 位元組（比 OpenAI 流行的 text-embedding-3-large 模型¹ 的未壓縮嵌入小 96 倍）實現高品質的檢索。就像 Arctic Embed 1.5 一樣，Arctic Embed 2.0 模型在壓縮狀態下也超越了一些支援 MRL 的同類產品，具有顯著較低的品質劣化和更高的基準分數。
真正的開源：Arctic Embed 2.0 模型是在寬鬆的 Apache 2.0 許可證下發布的。

Snowflake is excited to announce the release of Arctic Embed 2.0, the next iteration of our frontier embedding models, which now empower multilingual search. While our previous releases have been well received by our customers, partners and the open source community, leading to millions of downloads, we have consistently received one request: Can you make this model multilingual? Arctic Embed 2.0 builds on the robust foundation of our previous releases, adding multilingual support without sacrificing English performance or scalability, to address the needs of an even broader user base that spans a wide range of languages and applications.

![Snowflake data](/assets/library/snowflake-arctic-embed2/0546501b-9897-4145-af38-1b352fafb89c)
Figure 1. Single-vector dense retrieval performance of open source multilingual embedding models with fewer than 1B parameters. Scores are average nDCG@10 on MTEB Retrieval and the subset of CLEF (ELRA, 2006) covering English, French, Spanish, Italian and German.

### The diverse and powerful feature set of Arctic Embed 2.0
1. **Enterprise-ready throughput and efficiency:** The Arctic Embed 2.0 models are built for large-scale enterprise demands. Even our “large” model weighs in well under 1B parameters and delivers fast, high-throughput embedding capabilities. Based on internal testing, it easily handles more than 100 documents per second (on average) on NVIDIA A10 GPUs and achieves sub-10ms query embedding latency, enabling practical deployment on budget-friendly hardware.
2. **Uncompromising quality for English and non-English retrieval:** Despite their compact sizes, both Arctic Embed 2.0 models achieve impressive NDCG@10 scores across a variety of English and non-English benchmark data sets, demonstrating a capability to generalize well even to languages not included in the training recipe. These impressive benchmark scores position Arctic Embed 2.0 as a leader among frontier retrieval models.
3. **Enabling scalable retrieval through Matryoshka Representation Learning (MRL):** The Arctic Embed 2.0 release includes the same quantization-friendly MRL functionality introduced in Arctic Embed 1.5, allowing users to reduce cost and optimize scale when performing searches over large data sets. With both model sizes, users can achieve high-quality retrieval with as few as 128 bytes per vector (96x smaller than uncompressed embeddings from OpenAI’s popular text-embedding-3-large model1). Just like Arctic Embed 1.5, the Arctic Embed 2.0 models also outshine several MRL-supporting peers with substantially lower quality degradation and higher benchmark scores in the compressed regime.
4. **Truly open source:** The Arctic Embed 2.0 models are released under the permissive Apache 2.0 license.

貼上、拖放或點擊上傳圖片 (.png, .jpeg, .jpg, .svg, .gif)