selfgpt

selfgpt.pages.devDR 93 Accessible 300 views

Development & Programming An experimental project that runs chat models in reverse, letting them generate the user's side of a conversation while a judge model tries to distinguish their messages from real human ones.

Visit now Favorite

Editor's Review

For reference only

The site turns the usual chat setup inside out by exploiting chat templates rather than fine-tuning, and it compares a native reversed-template mode against simple roleplay prompting. Its automated judge ranking of models makes it a niche but genuinely interesting read for people curious about LLM b

Site introduction

selfgpt is a small experimental project built around a single question: can large language models be used in reverse? Instead of generating assistant replies, the models are asked to produce the user’s messages, flipping the usual direction of a chat conversation.

The core trick relies on chat templates. Normally a model generates after an assistant token and stops when it is the user’s turn. selfgpt applies the template backwards, letting the model continue past that stopping point so it writes the user message itself. A second configuration, roleplay, simply asks the model to act human, and the two approaches are compared side by side.

To evaluate the results, a judge model reviews the conversation and tries to tell model-written user messages apart from real people’s. This produces a ranking of which models are best at playing the user role in native mode, with notes on the judge’s reliability.

The project is aimed at people interested in LLM internals, prompt engineering, and evaluation methodology rather than at general chat users. It is a demonstration and write-up, not a hosted product, so visitors should expect to read the explanation and follow the linked tests rather than sign up for an account.