AB
AB
AiBoss

All Categories

project

harrier-oss-v1 - Microsoft's open-source multilingual text embedding model

Harrier-OSS-V1 is an open-source multilingual text embedding model from Microsoft, achieving state-of-the-art (SOTA) performance on the Multilingual MTEB v2 benchmark. The model employs a decoder-only architecture, utilizing final token pooling and L2 regression...

What is harrier-oss-v1?

Harrier-oss-v1 is an open-source multilingual text embedding model from Microsoft, achieving state-of-the-art (SOTA) performance on the Multilingual MTEB v2 benchmark. The model employs a decoder-only architecture, generating dense vectors through final token pooling and L2 normalization, making it suitable for tasks such as retrieval, clustering, semantic similarity, and classification. Harrier-oss-v1 is available in three versions: 27B, 0.6B, and 270M, balancing extreme performance with edge deployment requirements, and is open-source and commercially viable.

Main features of harrier-oss-v1

  • Text embeddingTransform the input text into a standardized high-dimensional dense vector representation using a decoder-only architecture.
  • Semantic retrievalEfficient document search and information retrieval based on vector similarity.
  • Text clusteringAutomatically group and categorize related texts based on semantic vectors.
  • Similarity calculationIt supports quantitative evaluation of the semantic relationship between two text segments.
  • Text classificationAutomatically classify text content using semantic feature vectors.
  • Bilingual miningSupports semantic alignment and matching retrieval of cross-language text.
  • The result was rearranged.: Optimize and sort candidate results based on semantic relevance to improve accuracy.

How to use harrier-oss-v1

  • Get the modelVisit the HuggingFace page to download version 27B, 0.6B, or 270M.
  • View documentRead the model cards and usage examples to understand the specific calling methods.
  • Loading ModelLoad pre-trained weights using HuggingFace Transformers or other frameworks.
  • Input text: Input the text sequence to be processed into the model for encoding.
  • Extract vector: Get the dense vector after the final token pooling and L2 normalization.
  • Downstream applications: Use vectors for specific business scenarios such as retrieval, clustering, and classification.

Project address for harrier-oss-v1

  • HuggingFace model library:
    • https://huggingface.co/microsoft/harrier-oss-v1-27b
    • https://huggingface.co/microsoft/harrier-oss-v1-0.6b
    • https://huggingface.co/microsoft/harrier-oss-v1-270m

Key information and usage requirements for harrier-oss-v1

  • ProducerMicrosoft
  • Model typeMultilingual text embedding model
  • ArchitectureDecoder-only
  • Core technologiesFinally, token pooling + L2 normalization.
  • Evaluation resultsMultilingual MTEB v2 SOTA (State-of-the-art)
  • Open source licenseFriendly open-source license (presumably MIT/Apache 2.0)
  • Model versionAvailable in three specifications: 27B / 0.6B / 270M.
  • Hardware RequirementsChoose the computing power according to the version; 27B requires sufficient GPU resources, while 270M can be deployed at the edge.
  • Software DependencyRequires HuggingFace Transformers or a compatible framework.
  • Input formatPlain text sequence

The core advantages of harrier-oss-v1

  • Leading performanceAchieved state-of-the-art (SOTA) results in the Multilingual MTEB v2 multilingual benchmark.
  • High-efficiency architectureThe model employs a decoder-only design combined with final token pooling to generate high-quality semantic vectors.
  • Flexible in scaleIt offers three versions: 27B, 0.6B, and 270M, covering all scenarios from high performance in the cloud to low power consumption at the edge.
  • Ready to use right out of the boxThe model is hosted on HuggingFace and can be downloaded and loaded directly without complex environment configuration.
  • A wide range of tasksA single model can simultaneously support six major tasks: retrieval, clustering, classification, similarity calculation, bilingual mining, and re-ranking.

Comparison of Harrier-OSS-V1 with similar competing products

Dimension Harrier-OSS-V1 E5-mistral-7b-instruct BGE-M3
Producer Microsoft Microsoft BAAI (智源)
Architecture Decoder-only Encoder-decoder Encoder (XLM-RoBERTa)
Core technologies Finally, token pooling + L2 normalization. Weakly supervised contrastive pre-training + instruction fine-tuning Multilingual, multi-function training + hybrid retrieval
Largest 27B 7B 9B (Multiple Sizes)
Lightweight version 0.6B, 270M none 568M
MTEB scores Multilingual MTEB v2 SOTA It dominated the charts for a long time (beyond Harrier's dominance). Leading in multilingual capabilities, and optimal for some tasks.
Output method Dense vectors Dense vectors Dense + Sparse + Multi-vector Mixture

Application scenarios of harrier-oss-v1

  • Semantic searchEfficient information retrieval and accurate search of large-scale document databases based on vector similarity.
  • Text clusteringBased on semantic vector features, it automatically and intelligently groups and categorizes massive amounts of text according to their topic relevance.
  • Intelligent classificationThe model uses text embedding feature vectors to automatically classify and categorize content types.
  • Semantic matchingIt enables precise matching scenarios such as text deduplication and content recommendation by calculating vector similarity.
  • Cross-language searchIt supports semantic alignment for dozens of languages, enabling cross-language text interoperability and retrieval.
Back