Show HN: Using an LLM chat template in reverse to play the user
Category: ai-ml
Tags: llm, chat-template, reverse-roleplay, evaluation, gguf
Score: 6.3/10 (Innovation: 7, Technical: 6, Documentation: 7, Utility: 5)
selfgpt is a Python tool that inverts LLM chat templates to make models generate user turns instead of assistant turns, enabling AI-to-AI conversation or role reversal. It includes an LLM-as-a-judge evaluation pipeline to measure how human-like models are when writing in the user slot, with results showing no model is indistinguishable from humans. The project is interesting for its creative exploitation of chat templates and its empirical study of model behavior.
Target audience: AI researchers, ML engineers, LLM hobbyists
Repository: https://selfgpt.pages.dev/ · Python · MIT
View on Hacker News