GRAG-BGE-M3-MERGED-x-SNOWFLAKE-ARCTIC-HESSIAN-AI

This is a merged sentence-transformers model. 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. Our GRAG-BGE-M3-MERGED Model was merged with Snowflake/snowflake-arctic-embed-l-v2.0 to exceed performances from each Base-Model.

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 the Snowflake Arctic Model (Snowflake), our Merged Model (Merged-BGE) and our Merged-BGE Model merged with Snowflake/snowflake-arctic-embed-l-v2.0

TASK	Snowflake	Merged-BGE	Merged-Snowflake	GRAG vs. Snowflake	Merged-Snowflake vs. Snowflake	Merged-Snowflake vs. Merged-BGE
AmazonCounterfactualClassification	0.6587	0.7111	0.7152	5.24%	5.65%	0.41%
AmazonReviewsClassification	0.3697	0.4571	0.4577	8.74%	8.80%	0.06%
FalseFriendsGermanEnglish	0.5360	0.5338	0.5378	-0.22%	0.18%	0.40%
GermanQuAD-Retrieval	0.9423	0.9311	0.9456	-1.12%	0.33%	1.45%
GermanSTSBenchmark	0.7499	0.8218	0.8558	7.19%	10.59%	3.40%
MassiveIntentClassification	0.6778	0.6522	0.6826	-2.56%	0.48%	3.04%
MassiveScenarioClassification	0.7375	0.7381	0.7494	0.06%	1.19%	1.13%
GermanDPR	0.8367	0.8159	0.8330	-2.08%	-0.37%	1.71%
MTOPDomainClassification	0.9080	0.9139	0.9259	0.59%	1.79%	1.20%
MTOPIntentClassification	0.6675	0.6684	0.7143	0.09%	4.68%	4.59%
PawsXPairClassification	0.5887	0.5710	0.5803	-1.77%	-0.84%	0.93%

Comparison between Original Base-Model (BGE-M3), Merged Model with Base-Model (Merged-BGE) and our Merged-BGE 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 GRAG-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
GRAG-BGE-M3-TRIPLES-HESSIAN-AI	0.8857
GRAG-BGE-M3-TRIPLES-MERGED-HESSIAN-AI	0.8866
GRAG-BGE-M3-MERGED-x-SNOWFLAKE-ARCTIC-HESSIAN-AI	0.8866
GRAG-UAE-LARGE-V1-TRIPLES-HESSIAN-AI	0.8763
GRAG-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.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("avemio/GRAG-BGE-M3-MERGED-x-SNOWFLAKE-ARCTIC-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}
}