Rudialogpt Small

🚀 Text Generation Model

This generation model is designed to build generative conversational agents, leveraging a large dialog dataset.

🚀 Quick Start

This generation model is based on sberbank-ai/rugpt3small_based_on_gpt2. It's trained on a large corpus of dialog data and can be used for building generative conversational agents.

The model was trained with a context size of 3.

✨ Features

Evaluation Metrics

On a private validation set, we calculated the metrics introduced in this paper:

• Sensibleness: Crowdsourcers were asked whether the model's response makes sense given the context.
• Specificity: Crowdsourcers were asked whether the model's response is specific for the given context. In other words, we don't want our model to give general and boring responses.
• SSA (Sensibleness Specificity Average): The average of the above two metrics.

💻 Usage Examples

Basic Usage

import torch
from transformers import AutoTokenizer, AutoModelWithLMHead

tokenizer = AutoTokenizer.from_pretrained('tinkoff-ai/ruDialoGPT-small')
model = AutoModelWithLMHead.from_pretrained('tinkoff-ai/ruDialoGPT-small')
inputs = tokenizer('@@ПЕРВЫЙ@@ привет @@ВТОРОЙ@@ привет @@ПЕРВЫЙ@@ как дела? @@ВТОРОЙ@@', return_tensors='pt')
generated_token_ids = model.generate(
    **inputs,
    top_k=10,
    top_p=0.95,
    num_beams=3,
    num_return_sequences=3,
    do_sample=True,
    no_repeat_ngram_size=2,
    temperature=1.2,
    repetition_penalty=1.2,
    length_penalty=1.0,
    eos_token_id=50257,
    max_new_tokens=40
)
context_with_response = [tokenizer.decode(sample_token_ids) for sample_token_ids in generated_token_ids]
context_with_response

📄 License

This project is licensed under the MIT license.