News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
DialoGPT

DialoGPT

A GPT-style model trained on Reddit threads to write chat replies

DialoGPT is a generative conversational model built by Microsoft researchers and described in a paper first posted in November 2019. It was trained on 147 million exchanges pulled from Reddit and is meant to produce replies that read like a human's in short chat exchanges.

Share on WhatsApp

  • Trained on 147 million Reddit exchanges from 2005 through 2017
  • Built by extending the Hugging Face PyTorch transformer
  • First posted on arXiv on 1 November 2019
  • Accepted as a system demonstration at ACL 2020
  • Weights and training pipeline released publicly

What it is

DialoGPT, short for dialogue generative pre-trained transformer, is a neural model that writes conversational replies. A Microsoft team introduced it in a paper dated 1 November 2019, later revised in 2020 and accepted for the ACL 2020 system demonstration track. It is a tunable response generator aimed at open-domain chat, meaning it is not limited to one topic. The authors trained it on 147 million conversation-like exchanges taken from Reddit comment chains covering 2005 through 2017. Their claim is that in single-turn exchanges it approaches human quality on both automatic and human-judged measures, and that its replies beat strong baselines on relevance, content and consistency with context.

How it came about

The work was carried out at Microsoft and written up by a nine-person research group. The first arXiv version appeared on 1 November 2019, with a second revision in April 2020 and the version shown on the listing dated 2 May 2020. The paper was accepted as a system demonstration at ACL 2020. Alongside the write-up, the team put out the trained weights and the training pipeline so that other researchers could build on the work. A model card on Hugging Face hosts one size of the model and points readers to the project repository for preprocessing and training details.

Inside the model

The model extends the Hugging Face PyTorch transformer implementation, so it follows the standard decoder-style transformer design used for text generation. Its training corpus is 147 million multi-turn dialogue exchanges drawn from Reddit discussion threads between 2005 and 2017. The paper describes it as large and tunable, meaning it can be adapted to particular tasks. The sample code on the model card shows a chat setup in which each new user turn is tokenised, appended to the running history and passed to the model, which then produces the next reply. In that example the whole conversation history is capped at 1000 tokens.

What it can do

The authors report that, in single-turn settings, the model's replies were judged close to human quality in both automatic scoring and human evaluation. They also state that systems built on it produced replies that were more relevant, fuller in content and more consistent with the surrounding context than strong baselines. The model card repeats the human evaluation claim, describing the output as comparable to a human's under a single-turn Turing test. The sample exchanges show multi-turn use, with the bot carrying on a short conversation across several turns.

How to use it

The pre-trained weights and the training pipeline were released publicly to support further research. A large version of the model is hosted on Hugging Face under the microsoft/DialoGPT-large identifier and can be loaded with the Transformers library as a causal language model. The model card includes runnable PyTorch code for a five-line chat. Preprocessing and training instructions live in the project repository on GitHub.

Why it matters

By publishing both the trained model and the pipeline behind it, the team aimed to give other researchers a starting point for work on neural response generation and on more capable open-domain dialogue systems. The paper was accepted as a system demonstration at ACL 2020, and the weights were made available through Hugging Face for reuse.

Questions people ask about DialoGPT

What is DialoGPT for?

It is a text generation model that writes replies in a conversation. It was built for open-domain chat, so it can respond on any topic rather than one narrow subject, and the authors tested it mainly on single-turn exchanges.

Who made it and when?

A Microsoft research group built it. The paper was first posted on arXiv on 1 November 2019, revised in April and May 2020, and accepted for the ACL 2020 system demonstration track.

What was it trained on?

The training data is 147 million conversation-like exchanges taken from Reddit comment chains spanning 2005 to 2017. The model card describes the same figure as multi-turn dialogue from Reddit discussion threads.

How good are its replies?

The authors say that in single-turn settings the output came close to human quality on both automatic measures and human judgements, and that it beat strong baselines on relevance, content and context consistency. The model card repeats the human evaluation claim.

Can I use it myself?

Yes. The weights and the training pipeline were released publicly, and a large version is hosted on Hugging Face as microsoft/DialoGPT-large, loadable with the Transformers library. The model card includes sample code for a short chat.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on DialoGPT yet, so the facts on this page come from the sources listed above: arxiv.org/abs/1911.00536, huggingface.co/microsoft/DialoGPT-large. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.