gpt2bot  by polakowo

Telegram chatbot using transformers for multi-turn dialogue

created 5 years ago
439 stars

Top 69.1% on sourcepulse

GitHubView on GitHub
Project Summary

This project provides a Telegram chatbot powered by large language models, specifically DialoGPT, for engaging multi-turn conversations. It's designed for users interested in deploying AI-powered conversational agents, offering a fun, community-like interaction style due to its Reddit-based training data.

How It Works

The chatbot leverages Hugging Face's transformers library to run pre-trained dialogue generation models like DialoGPT. DialoGPT, trained on extensive Reddit dialogue, provides human-comparable response quality in single-turn Turing tests. For enhanced dialogue quality, DialogRPT models, trained on human feedback, are employed for response ranking. Users can also integrate other text generators supported by transformers.

Quick Start & Requirements

  • Install via pip install -r requirements.txt after cloning the repository.
  • Requires Python 3.6+ and a Telegram Bot API token. A GIPHY API key is needed for GIF generation.
  • Configuration files (.cfg) specify model size (e.g., medium-cpu, large-gpu) and ranking strategies.
  • Console testing is available via python run_bot.py --type=console.

Highlighted Details

  • Utilizes DialoGPT, a large-scale pretrained dialogue response generation model.
  • Incorporates DialogRPT for improved dialogue quality through response ranking.
  • Supports GIF generation via GIPHY API integration.
  • Offers multiple configuration options for CPU and GPU, with varying ranking models.

Maintenance & Community

No specific contributor or community links (Discord, Slack, etc.) are mentioned in the README.

Licensing & Compatibility

The README does not explicitly state the license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The README suggests manual tuning of parameters like temperature, top_k, and top_p is necessary to achieve desired conversational styles, implying a trial-and-error process for optimal performance. The model's Reddit training data can lead to unpredictable or "off-topic" responses, especially with higher temperature settings.

Health Check
Last commit

1 year ago

Responsiveness

1 day

Pull Requests (30d)
0
Issues (30d)
0
Star History
2 stars in the last 90 days

Explore Similar Projects

Feedback? Help us improve.