SLX-v0.1

SLX-v0.1 is an advanced model developed by BRAHMAI Research, specifically designed for mapping sentences and paragraphs into a 384-dimensional dense vector space. This model is ideal for tasks such as clustering, semantic search, and sentence similarity.

Usage with Sentence-Transformers

Using SLX-v0.1 is straightforward with the Sentence-Transformers library. Follow the instructions below to get started:

Installation

pip install -U sentence-transformers

Example

from sentence_transformers import SentenceTransformer

sentences = ["This is an example sentence", "Each sentence is converted"]
model = SentenceTransformer('brahmairesearch/slx-v0.1')
embeddings = model.encode(sentences)

print(embeddings)

Usage with Hugging Face Transformers

If you prefer to use Hugging Face Transformers without the Sentence-Transformers library, you can still utilize the SLX-v0.1 model. Below is a guide on how to process input and apply mean pooling to obtain sentence embeddings:

from transformers import AutoTokenizer, AutoModel
import torch
import torch.nn.functional as F

# Mean Pooling - Takes attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0]  # Token embeddings from the model output
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

# Sentences for embedding
sentences = ['This is an example sentence', 'Each sentence is converted']

# Load model from Hugging Face Hub
tokenizer = AutoTokenizer.from_pretrained('brahmairesearch/slx-v0.1')
model = AutoModel.from_pretrained('brahmairesearch/slx-v0.1')

# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform pooling
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

# Normalize embeddings
sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)

print("Sentence embeddings:")
print(sentence_embeddings)

Intended Use Cases

SLX-v0.1 is designed for encoding sentences and short paragraphs into dense vectors that capture their semantic information. These embeddings can be used for:

• Information retrieval
• Clustering
• Sentence similarity tasks

Please note that input text longer than 256 word pieces will be truncated by default.

Training Procedure

Pre-training

SLX-v0.1 is based on the pre-trained model nreimers/MiniLM-L6-H384-uncased. For more information on the pre-training process, please refer to the original model card.

Fine-tuning

The model was fine-tuned using a contrastive learning objective. Cosine similarity is computed between all possible sentence pairs within a batch, and cross-entropy loss is applied using the true pairs.

Following fine-tuning, the model underwent transfer learning using the dunzhang/stella_en_400M_v5 model with an internally curated dataset, optimizing it for English language tasks.

Contact

For inquiries or feedback, feel free to reach out to us at [email protected].

Best regards,
The BRAHMAI Team