Model Summary:
Granite-Embedding-30m-English is a 30M parameter dense bi-encoder embedding model from the Granite Embeddings suite that can be used to generate high quality text embeddings. This model produces embedding vectors of size 384 and is trained using a combination of open source relevance-pair datasets with permissive, enterprise-friendly license, and IBM collected and generated datasets. While maintaining competitive scores on academic benchmarks such as BEIR, this model also performs well on many enterprise use cases. This model is developed using retrieval oriented pre-training, contrastive fine-tuning, knowledge distillation and model merging for improved performance.
Granite-embedding-30m-r1.1 was specifically designed to support multi-turn information retrieval and is designed to handle contextual document retrieval in multi-turn conversational information retrieval. Granite-embedding-30m-r1.1 was trained on data tailored for multi-turn conversational information retrieval and uses multi-teacher distillation over granite-embedding-30m-english (https://huggingface.co/ibm-granite/granite-embedding-30m-english)
Intended use:
The model is designed to produce fixed length vector representations for a given text, which can be used for text similarity, retrieval, and search applications.
Usage with Transformers.js:
This is a simple example of how to use the granite-embedding-small-english-r2 model with the Transformers.js library.
If you haven't already, you can install the Transformers.js JavaScript library from NPM using:
npm i @huggingface/transformers
The model can then be used to encode pairs of text
js
1import{AutoModel,AutoTokenizer, matmul }from"@huggingface/transformers";23// Download from the 🤗 Hub4const model_id ="onnx-community/granite-embedding-30m-english-ONNX";5const tokenizer =awaitAutoTokenizer.from_pretrained(model_id);6const model =awaitAutoModel.from_pretrained(model_id,{7dtype:"fp32",// Options: "fp32" | "fp16" | "q8" | "q4" | "q4f16"8});910// Prepare queries and documents11const input_queries =[12" Who made the song My achy breaky heart? ",13"summit define",14];15const input_passages =[16"Achy Breaky Heart is a country song written by Don Von Tress. Originally titled Don't Tell My Heart and performed by The Marcy Brothers in 1991. ",17"Definition of summit for English Language Learners. : 1 the highest point of a mountain : the top of a mountain. : 2 the highest level. : 3 a meeting or series of meetings between the leaders of two or more governments.",18];19const inputs =awaittokenizer([...input_queries,...input_passages],{20padding:true,21});2223// Generate embeddings24const{ sentence_embedding }=awaitmodel(inputs);25const normalized_sentence_embedding = sentence_embedding.normalize();2627// Compute similarities28const scores =awaitmatmul(29 normalized_sentence_embedding.slice([0, input_queries.length]),30 normalized_sentence_embedding
31.slice([input_queries.length,null])32.transpose(1,0),33);34const scores_list = scores.tolist();35console.log(scores_list);36// [37// [ 0.84034663438797, 0.498334139585495 ],38// [ 0.4650711715221405, 0.5818834900856018 ]39// ]
Evaluation:
Granite-Embedding-30M-English model is twice as fast as other models with similar embedding dimensions, while maintaining competitive performance. The performance of the Granite-Embedding-30M-English model on MTEB Retrieval (i.e., BEIR) and code retrieval (CoIR) benchmarks is reported below.
Model
Paramters (M)
Embedding Dimension
MTEB Retrieval (15)
CoIR (10)
granite-embedding-30m-english
30
384
49.1
47.0
granite-embedding-30m-r1.1 revision maintains the fast speed of granite-embedding-30m-english while demontratring strong performance on multi-turn information retrieval benchmarks. The performance of the granite-embedding-30M-r1.1 model on MTEB Retrieval (i.e., BEIR) and multi-turn information retrieval (MTRAG(https://github.com/IBM/mt-rag-benchmark), Multidoc2dial(https://github.com/IBM/multidoc2dial)) datasets is reported below.
Model
Parameters (M)
Embedding Dimension
MTEB Retrieval (15)
MT-RAG
Mdoc2dial
granite-embedding-30m-english
30
384
49.1
49.16
85.42
granite-embedding-30m-english-r1.1
30
384
48.9
52.33
85.78
bge-small-en-v1.5
33
512
53.86
38.26
83.71
e5-small-v2
33
384
48.46
28.72
75.7
Model Architecture:
granite-embedding-30m-English is based on an encoder-only RoBERTa like transformer architecture, trained internally at IBM Research. granite-embedding-30m-r1.1 shares the same architecture as granite-embedding-30m-English
Model
granite-embedding-30m-english
granite-embedding-125m-english
granite-embedding-107m-multilingual
granite-embedding-278m-multilingual
Embedding size
384
768
384
768
Number of layers
6
12
6
12
Number of attention heads
12
12
12
12
Intermediate size
1536
3072
1536
3072
Activation Function
GeLU
GeLU
GeLU
GeLU
Vocabulary Size
50265
50265
250002
250002
Max. Sequence Length
512
512
512
512
# Parameters
30M
125M
107M
278M
Training Data:
Overall, the training data consists of four key sources: (1) unsupervised title-body paired data scraped from the web, (2) publicly available paired with permissive, enterprise-friendly license, (3) IBM-internal paired data targeting specific technical domains, and (4) IBM-generated synthetic data. The data is listed below:
Notably, we do not use the popular MS-MARCO retrieval dataset in our training corpus due to its non-commercial license, while other open-source models train on this dataset due to its high quality.
Infrastructure:
We train Granite Embedding Models using IBM's computing cluster, Cognitive Compute Cluster, which is outfitted with NVIDIA A100 80gb GPUs. This cluster provides a scalable and efficient infrastructure for training our models over multiple GPUs.
Ethical Considerations and Limitations:
The data used to train the base language model was filtered to remove text containing hate, abuse, and profanity. Granite-embedding-30m-english and Granite-embedding-30m-r1.1 are trained only for English texts, and has a context length of 512 tokens (longer texts will be truncated to this size).
