metadata

pipeline_tag: sentence-similarity
tags:
  - sentence-transformers
  - feature-extraction
  - sentence-similarity
license: apache-2.0
datasets:
  - wikimedia/wikipedia
  - SiberiaSoft/SiberianPersonaChat-2
language:
  - ru
  - en
metrics:
  - mse
library_name: transformers

FractalGPT/SberDistil

This is a sentence-transformers model: It maps sentences & paragraphs to a 384 dimensional dense vector space and can be used for tasks like clustering or semantic search. This is a fast and small model for solving the problem of determining the proximity between sentences, in the future we will reduce and speed it up. Project

Usage (Sentence-Transformers)

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers

Then you can use the model like this:

import numpy as np
from sentence_transformers import SentenceTransformer

model = SentenceTransformer('FractalGPT/SberDistil')

def cos(x, y):
  return np.dot(x, y)/(np.linalg.norm(x)*np.linalg.norm(y))

text_1 = "Кто такой большой кот?"
text_2 = "Who is kitty?"
a = model.encode(text_1)
b = model.encode(text_2)
cos(a, b)

>>> 0.8072159157330788

Training

The original weights was taken from cointegrated/rubert-tiny2.
Training was conducted in two stages:

In the first stage, the model was trained on Wikipedia texts (4 million texts) for three epochs.
In the second stage, training was conducted on Wikipedia, a dialog dataset, and NLI for one epoch.

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 312, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
  (2): Dense({'in_features': 312, 'out_features': 384, 'bias': True, 'activation_function': 'torch.nn.modules.linear.Identity'})
)