Most datasets teach models to produce answers; Antidoom Mix v1.0 instead supplies only prompts that provoke challenging generative behavior so you can mine and train targeted preference pairs. That focus makes it practical to generate FTPO (Final Token Preference Optimization) training examples that teach a model to avoid single-token loop-starts without adding new task-specific gold answers.
What Sets It Apart
- Prompt-only composition: every row intentionally excludes gold answers, rationales, verifier targets and terminal answer cues, so downstream pipelines can generate model completions and mine loop failures without label leakage. This means it is optimized for generation-and-mine workflows rather than supervised fine-tuning.
- Broad, mixed sources: the mixture bundles ~478,229 final rows drawn from many permissively licensed public splits (math, MMLU auxiliary, ShareGPT-style conversational prompts, code/apps etc.), giving diverse failure modes that surface repetition during reasoning and coding.
- Built for FTPO/Antidoom workflows: designed to be consumed by Antidoom-style pipelines that generate completions, detect loop-start tokens, and create chosen/rejected token pairs for targeted preference training (e.g., LoRA adapters that mitigate doom loops).
Who it's for and tradeoffs
Great fit if you want to: mine model failures and build preference-training data to reduce repetitive degeneration; evaluate model looping behavior across math, coding and multi-step reasoning prompts; or bootstrap FTPO-style fine-tuning without assembling prompts yourself. Look elsewhere if you need: labeled gold answers, evaluation/test splits, or turnkey supervised datasets — Antidoom Mix deliberately omits those. Also note the mixture contains rows sourced from multiple third-party datasets and the card lists license provenance; users must verify downstream licensing for redistribution or commercial use.
Where it fits
This dataset is a component in data-centric pipelines that aim to remove a narrow but impactful failure mode (doom loops) rather than to teach new task knowledge. It pairs naturally with Antidoom generation-and-train code and preference-optimization trainers that target single-token preferences, and is complementary to supervised instruction or RLHF datasets when the goal is behavioral repair rather than capability expansion.
Practical notes
- Size: ~478,229 final rows (prompt-only).
- Licensing: published under an Apache-2.0 / mixed-permissive notice on the HF card; individual sources and provenance are enumerated and should be checked for downstream use.
- Limitations: not a drop-in supervised QA/train set (no answers/labels), and using it requires a generation-and-mining pipeline to produce training pairs.
