Summary

Antidoom addresses "doom loops"—repetitive text generation in LLMs—by generating targeted preference data and training LoRA adapters. It's a specialized tool for improving LLM output coherence and reducing undesirable repetition, particularly during reasoning tasks.

How It Works

The project employs Final Token Preference Optimization (FTPO) to combat repetition. It generates completions, detects repetitive spans, marks the first loop-starting token as rejected, and selects coherent alternatives. This local preference data trains a LoRA adapter, teaching the model to avoid initiating repetition loops. This approach uniquely targets the precise token where repetition begins, adapting Antislop's methodology.

Quick Start & Requirements

• Installation: Clone repo (git clone https://github.com/Liquid4All/antidoom), cd antidoom, uv sync.
• Prerequisites: Supported GPU (NVIDIA/CUDA default; AMD/ROCm requires separate ROCm venv with ROCm 6.3+/7.0+ and Python 3.12). PyTorch and vLLM are necessary.
• Configuration: Edit configs/default.yaml for model_name. Default pipeline reads prompts from LiquidAI/antidoom-mix-v1.0, generates FTPO pairs, trains LoRA, and merges.
• Run Command: uv run antidoom -c configs/default.yaml -r runs/antidoom1 --temp 0.01 --model-name LiquidAI/LFM2.5-1.2B-Base.
• Links: Repo, Dataset.

Highlighted Details

• FTPO trains on local token preferences, targeting the first token of a repetition loop.
• Default dataset LiquidAI/antidoom-mix-v1.0 is a prompt-only ShareGPT mixture.
• Detailed AMD/ROCm support includes specific configurations (attention_backend: TRITON_ATTN, optim: adamw_torch) and environment setup.
• Extensive hyperparameter guidance for tuning max_train_examples, learning_rate, and early_stopping_chosen_win to prevent overtraining and ensure regularization.

Maintenance & Community

• Maintained by "Liquid AI".
• No community channels (Discord, Slack) or social media links provided.
• No roadmap detailed.

Licensing & Compatibility

• License type is not explicitly stated.
• Commercial use compatibility is not specified.

Limitations & Caveats

• Targets a "narrow failure mode" (model repetition loops); may not address other generation issues.
• AMD/ROCm setup is complex, requiring specific environment management and configuration overrides.
• Effective use requires careful hyperparameter tuning to avoid model degradation or insufficient loop reduction.
• Absence of a stated license is a significant adoption blocker.