Antidoom Mix v1.0 is a pure cue word dataset released by Liquid AI in 2026. It aims to provide pure cue word training data for antidoom generation (anti-doom loop generation) and preference data construction pipelines. By inducing the model to expose doom loop behavior and constructing preferences to train the model, it achieves anti-doom loop generation, thereby improving the generation reliability and termination ability of the inference model. This dataset contains 478,229 lines of prompts, sourced from 14 upstream open-source datasets, covering task types such as mathematical reasoning, common sense question answering, academic research, programming code, instruction fine-tuning, structured output, and preference data. Each data point retains only the user's question, excluding standard answers, reasoning processes, hidden tests or verification targets, and removes end-answer prompts and lines that clearly contain answer trajectories.

Citation

@article{liquidAI2026Antidoom,
author = {Liquid AI},
title = {Reducing Doom Loops with Final Token Preference Optimization},
journal = {Liquid AI Blog},
year = {2026},
note = {www.liquid.ai/blog/antidoom}
}