wtf is antidoom?

Antidoom is a tool that fixes a specific problem some AI language models run into called a doom loop, where the model gets stuck repeating the same words or phrases over and over instead of continuing to reason toward an answer. This tends to happen with certain reasoning words like Wait, So, But, or Alternatively, which can become overused after heavy training, especially when the model is set to produce very predictable, low-randomness output and once it starts repeating itself there is nothing pulling it back on track. Instead of retraining a model on entire correct answers, Antidoom works at the exact point where a repeat begins. It generates sample answers from a model, scans them to find where a repeated section starts, and marks that specific word as the wrong choice. It then picks other reasonable words the model could have chosen instead at that same spot. These paired examples, one rejected word and one or more better alternatives, become training data. That training data is used to train a small add-on called a LoRA adapter, using a method called Final Token Preference Optimization, which nudges the model away from the specific words that tend to trigger repetition, without changing everything else about how the model behaves. The result is meant to be a model that loops less often while still reasoning normally. To use it, someone clones the repository, installs dependencies, and edits a configuration file to point at the model checkpoint they want to fix. Running the tool then generates example completions, builds the training pairs, trains the adapter, and merges it into the model, all inside one command. It relies on a GPU, with the default setup built for NVIDIA and CUDA, and a separate setup path provided for AMD's ROCm hardware. The project ships with a companion prompt dataset built specifically for this pipeline, and a configuration file with settings a user can adjust, such as how many training examples to use and how strongly to prevent overused words from dominating the results. The README includes practical starting values for these settings based on the maintainers' own experiments.