This is a sentence-transformers model trained on this Dataset with roughly 300k Triple-Samples. It maps sentences & paragraphs to a 1024-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more. It was merged with the Base-Model BAAI/bge-m3 again to maintain performance on other languages again.

Model Details

Model Description

Model Type: Sentence Transformer

Maximum Sequence Length: 8192 tokens
Output Dimensionality: 1024 tokens
Similarity Function: Cosine Similarity

Model Sources

Documentation: Sentence Transformers Documentation
Repository: Sentence Transformers on GitHub
Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False}) with Transformer model: XLMRobertaModel 
  (1): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): Normalize()
)

Evaluation MTEB-Tasks

Classification

AmazonCounterfactualClassification
AmazonReviewsClassification
MassiveIntentClassification
MassiveScenarioClassification
MTOPDomainClassification
MTOPIntentClassification

Pair Classification

FalseFriendsGermanEnglish
PawsXPairClassification

Retrieval

GermanQuAD-Retrieval
GermanDPR

STS (Semantic Textual Similarity)

GermanSTSBenchmark

Comparison between Base-Model (BGE-M3), Finetuned Model (German-RAG-BGE) and Merged Model with Base-Model (Merged-BGE)

TASK	BGE-M3	German-RAG-BGE	Merged-BGE	German-RAG vs. BGE	Merged vs. BGE
AmazonCounterfactualClassification	0.6908	0.5449	0.7111	-14.59%	2.03%
AmazonReviewsClassification	0.4634	0.2745	0.4571	-18.89%	-0.63%
FalseFriendsGermanEnglish	0.5343	0.4777	0.5338	-5.67%	-0.05%
GermanQuAD-Retrieval	0.9444	0.8714	0.9311	-7.30%	-1.33%
GermanSTSBenchmark	0.8079	0.7921	0.8218	-1.58%	1.39%
MassiveIntentClassification	0.6575	0.4884	0.6522	-16.90%	-0.52%
MassiveScenarioClassification	0.7355	0.5837	0.7381	-15.19%	0.25%
GermanDPR	0.8265	0.7210	0.8159	-10.54%	-1.06%
MTOPDomainClassification	0.9121	0.7450	0.9139	-16.71%	0.17%
MTOPIntentClassification	0.6808	0.4516	0.6684	-22.92%	-1.25%
PawsXPairClassification	0.5678	0.5077	0.5710	-6.01%	0.33%

Comparison between Base-Model (BGE-M3), Merged Model with Base-Model (Merged-BGE) and our Merged-Model merged with Snowflake/snowflake-arctic-embed-l-v2.0

TASK	BGE-M3	Merged-BGE	Merged-Snowflake	Merged-BGE vs. BGE	Merged-Snowflake vs. BGE	Merged-Snowflake vs. Merged-BGE
AmazonCounterfactualClassification	0.6908	0.7111	0.7152	2.94%	3.53%	0.58%
AmazonReviewsClassification	0.4634	0.4571	0.4577	-1.36%	-1.23%	0.13%
FalseFriendsGermanEnglish	0.5343	0.5338	0.5378	-0.09%	0.66%	0.75%
GermanQuAD-Retrieval	0.9444	0.9311	0.9456	-1.41%	0.13%	1.56%
GermanSTSBenchmark	0.8079	0.8218	0.8558	1.72%	5.93%	4.14%
MassiveIntentClassification	0.6575	0.6522	0.6826	-0.81%	3.82%	4.66%
MassiveScenarioClassification	0.7355	0.7381	0.7494	0.35%	1.89%	1.53%
GermanDPR	0.8265	0.8159	0.8330	-1.28%	0.79%	2.10%
MTOPDomainClassification	0.9121	0.9139	0.9259	0.20%	1.52%	1.31%
MTOPIntentClassification	0.6808	0.6684	0.7143	-1.82%	4.91%	6.87%
PawsXPairClassification	0.5678	0.5710	0.5803	0.56%	2.18%	1.63%

Evaluation on German-RAG-EMBEDDING-BENCHMARK

Accuracy is calculated by evaluating if the relevant context is the highest ranking embedding of the whole context array. See Eval-Dataset and Evaluation Code here

Model Name	Accuracy
bge-m3	0.8806
UAE-Large-V1	0.8393
German-RAG-BGE-M3-TRIPLES-HESSIAN-AI	0.8857
German-RAG-BGE-M3-TRIPLES-MERGED-HESSIAN-AI	0.8866
German-RAG-BGE-M3-MERGED-x-SNOWFLAKE-ARCTIC-HESSIAN-AI	0.8866
German-RAG-UAE-LARGE-V1-TRIPLES-HESSIAN-AI	0.8763
German-RAG-UAE-LARGE-V1-TRIPLES-MERGED-HESSIAN-AI	0.8771

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

python

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("avemio/German-RAG-BGE-M3-TRIPLES-HESSIAN-AI")
# Run inference
sentences = [
    'The weather is lovely today.',
    "It's so sunny outside!",
    'He drove to the stadium.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Training Details

Framework Versions

Python: 3.10.12
Sentence Transformers: 3.2.1
Transformers: 4.44.2
PyTorch: 2.4.1+cu121
Accelerate: 0.34.2
Datasets: 3.0.1
Tokenizers: 0.19.1

Citation

@misc{bge-m3,
      title={BGE M3-Embedding: Multi-Lingual, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation}, 
      author={Jianlv Chen and Shitao Xiao and Peitian Zhang and Kun Luo and Defu Lian and Zheng Liu},
      year={2024},
      eprint={2402.03216},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

The German-RAG AI Team

Marcel Rosiak Soumya Paul Siavash Mollaebrahim Zain ul Haq