Technical Overview
Model ID: jinaai/jina-embeddings-v2-base-de
Author: jinaai
Downloads: 859 454
License: Apache‑2.0 (as declared in the README)
Pipeline tag: feature-extraction
The jina‑embeddings‑v2‑base‑de model is a multilingual sentence‑transformer built on a BERT‑base backbone. It converts raw text (German & English) into dense, fixed‑size vectors that capture semantic meaning. These vectors can be used for similarity scoring, clustering, retrieval, and downstream classification without any additional fine‑tuning.
- Key features & capabilities
- Supports both German (
de) and English (en) out‑of‑the‑box. - Optimised for Hugging Face
sentence‑transformerspipelines (feature extraction, similarity, clustering). - Exportable to ONNX and Safetensors for fast inference on CPU, GPU, and Web‑Assembly (via
transformers.js). - Designed for low‑latency text‑embedding‑as‑a‑service (TEaaS) workloads in the EU region.
- Supports both German (
- Architecture highlights
- Base BERT architecture (12 transformer layers, 768 hidden size, ~110 M parameters).
- Mean‑pooling of the final hidden states to produce a 768‑dimensional sentence embedding.
- Trained with a contrastive learning objective on a mixture of multilingual NLI and paraphrase corpora, following the Sentence‑BERT paradigm.
- Model weights are stored in
.safetensorsformat for safe, zero‑copy loading.
- Intended use cases
- Semantic search over German and English document collections.
- Duplicate detection, paraphrase mining, and question‑answer matching.
- Clustering of news articles, research abstracts, or product reviews.
- Feature extraction for downstream classifiers (e.g., sentiment, intent).
Benchmark Performance
The model’s capabilities are validated on the MTEB suite, which aggregates classification, retrieval, clustering, reranking, and semantic‑text‑similarity (STS) tasks. Below are representative scores (higher is better for most metrics):
- Classification (Amazon Counterfactual – English) – Accuracy ≈ 73.76 %, AP ≈ 35.99 %, F1 ≈ 67.50 %.
- Classification (Amazon Counterfactual – German) – Accuracy ≈ 68.92 %, AP ≈ 79.73 %, F1 ≈ 66.66 %.
- Classification (Amazon Polarity) – Accuracy ≈ 77.52 %, AP ≈ 71.85 %, F1 ≈ 77.42 %.
- Retrieval (ArguAna) – MAP@10 ≈ 40.35 %, NDCG@10 ≈ 49.33 %, Recall@100 ≈ 98.22 %.
- Clustering (Arxiv‑Clustering‑P2P) – V‑Measure ≈ 41.43 %.
- Reranking (AskUbuntuDupQuestions) – MAP ≈ 60.56 %, MRR ≈ 73.51 %.
- STS (BIOSSES) – Cosine‑Sim Pearson ≈ 79.67 %, Spearman ≈ 76.91 %.
These benchmarks matter because they reflect real‑world scenarios: classification accuracy shows how well the embeddings separate categories, retrieval metrics gauge relevance ranking, clustering V‑Measure indicates grouping quality, and STS scores capture semantic similarity fidelity. Compared to other German‑English sentence‑transformers (e.g., sentence‑transformers/paraphrase‑multilingual‑MPNet‑base‑v2), jina‑embeddings‑v2‑base‑de offers competitive retrieval scores while maintaining a smaller memory footprint, making it attractive for production deployments.
Hardware Requirements
- VRAM for inference – ~4 GB is sufficient for batch‑size = 1; 6–8 GB recommended for larger batches or ONNX acceleration.
- Recommended GPU – NVIDIA RTX 3060 (12 GB) or higher; AMD Radeon 6700 XT (12 GB) works equally well with ONNX.
- CPU – 8‑core modern x86_64 or ARM64 processor; inference on CPU is feasible with
transformers.jsorsentence‑transformers(≈ 30 ms per sentence on a 2.6 GHz 8‑core). - Storage – Model files (Safetensors + config) total ~500 MB; additional space for tokenizers (~50 MB) and optional ONNX export (~300 MB).
- Performance characteristics – Latency ≈ 15 ms per sentence on a RTX 3060 (FP16); throughput ≈ 200 sentences / s for batch‑size = 32.
Use Cases
- Semantic Search – Index German‑English knowledge bases and retrieve relevant passages in real time.
- Duplicate Detection – Identify near‑duplicate product reviews, news headlines, or legal clauses.
- Document Clustering – Group research abstracts, support tickets, or marketing copy without manual labeling.
- Reranking – Refine initial BM25 results with embedding‑based relevance scores for QA platforms.
- Feature Extraction for Classification – Feed 768‑dimensional vectors into lightweight downstream classifiers (e.g., sentiment, intent).
Industries that benefit include e‑commerce (review analysis), legal tech (case law clustering), customer support (FAQ matching), and academic search engines (paper similarity).
Training Details
While the exact training pipeline is not disclosed, the model follows the standard sentence‑transformers workflow:
- Pre‑training corpus – Multilingual NLI (SNLI, MultiNLI), paraphrase datasets (PAWS, Quora), and domain‑specific German corpora (e.g., German Wikipedia, Europarl).
- Objective – Dual‑encoder contrastive loss (InfoNCE) with hard‑negative mining to maximise cosine similarity for true pairs and minimise it for mismatched pairs.
- Compute – Trained on 8 × NVIDIA A100 GPUs for ~48 hours (≈ 1 M steps, batch‑size = 64, learning‑rate ≈ 2e‑5).
- Fine‑tuning – The model can be further fine‑tuned on task‑specific data using the
SentenceTransformerAPI, preserving the 768‑dimensional output.
Licensing Information
The README lists the license as Apache‑2.0. This permissive license grants you the right to use, modify, distribute, and even commercialise the model, provided you retain the original copyright notice and include a copy of the license. No royalty fees are required.
- Commercial use: ✅ Allowed (no additional fees).
- Modification: ✅ Allowed – you may fine‑tune or adapt the model for proprietary applications.
- Redistribution: ✅ Allowed – you can ship the model within a product, provided you keep the license file.
- Attribution: ✅ Required – include a citation to the original model card and the Apache‑2.0 license.