README
Antidoom Mix v1.0
[!Note] 📝 Blog post: https://www.liquid.ai/blog/antidoom
💻 GitHub: https://github.com/Liquid4All/antidoom
Antidoom Mix v1.0 is a prompt-only training mixture for antidoom-style generation and preference-data pipelines. Responses are generated on this dataset, and looping traces are retained to construct preference pairs.
The dataset is intended to provide prompts only. Gold answers, rationales, hidden tests, verifier targets, and answer labels are intentionally removed.
Terminal answer cues such as Answer: are stripped, and rows with obvious answer traces are filtered out.
The initial release excludes public eval/test prompt sets and noncommercial or unclear redistribution sources.
Sources
| source_id | source | split/config | license | final rows |
|---|---|---|---|---|
gsm8k_train | openai/gsm8k | main/train | MIT | 7,473 |
math_lighteval_train | DigitalLearningGmbH/MATH-lighteval | 7 train configs | MIT | 7,499 |
mmlu_auxiliary_train | cais/mmlu | all/auxiliary_train | MIT | 50,000 |
math_qa_train | allenai/math_qa | default/train | Apache-2.0 | 29,836 |
commonsense_qa_train | tau/commonsense_qa | default/train | MIT | 9,741 |
pubmedqa_artificial_train | pubmed_qa | pqa_artificial/train | MIT | 50,000 |
apps_train | codeparrot/apps | train | MIT | 5,000 |
open_perfectblend_metamathqa | meta-math/MetaMathQA via mlabonne/open-perfectblend | train | MIT | 50,000 |
open_perfectblend_ultrainteract | openbmb/UltraInteract_sft via mlabonne/open-perfectblend | train | MIT | 49,540 |
open_perfectblend_ultrachat200k | mlabonne/ultrachat_200k_sft, from HuggingFaceH4/ultrachat_200k, via mlabonne/open-perfectblend | train | MIT | 49,958 |
open_perfectblend_evol_codealpaca | theblackcat102/evol-codealpaca-v1 via mlabonne/open-perfectblend | train | Apache-2.0 | 49,981 |
open_perfectblend_autoif | Post-training-Data-Flywheel/AutoIF-instruct-61k via mlabonne/open-perfectblend | train | Apache-2.0 | 49,995 |
open_perfectblend_lmsys_arena | mlabonne/lmsys-arena-human-preference-55k-sharegpt via mlabonne/open-perfectblend | train | Apache-2.0 | 49,206 |
ifstruct_train_generated | LiquidAI/ifstruct_generated | generated train rows | MIT | 20,000 |
Total final rows: 478,229.
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}
}
Similar repos
README
Antidoom Mix v1.0
[!Note] 📝 Blog post: https://www.liquid.ai/blog/antidoom
💻 GitHub: https://github.com/Liquid4All/antidoom
Antidoom Mix v1.0 is a prompt-only training mixture for antidoom-style generation and preference-data pipelines. Responses are generated on this dataset, and looping traces are retained to construct preference pairs.
The dataset is intended to provide prompts only. Gold answers, rationales, hidden tests, verifier targets, and answer labels are intentionally removed.
Terminal answer cues such as Answer: are stripped, and rows with obvious answer traces are filtered out.
The initial release excludes public eval/test prompt sets and noncommercial or unclear redistribution sources.
Sources
| source_id | source | split/config | license | final rows |
|---|---|---|---|---|
gsm8k_train | openai/gsm8k | main/train | MIT | 7,473 |
math_lighteval_train | DigitalLearningGmbH/MATH-lighteval | 7 train configs | MIT | 7,499 |
mmlu_auxiliary_train | cais/mmlu | all/auxiliary_train | MIT | 50,000 |
math_qa_train | allenai/math_qa | default/train | Apache-2.0 | 29,836 |
commonsense_qa_train | tau/commonsense_qa | default/train | MIT | 9,741 |
pubmedqa_artificial_train | pubmed_qa | pqa_artificial/train | MIT | 50,000 |
apps_train | codeparrot/apps | train | MIT | 5,000 |
open_perfectblend_metamathqa | meta-math/MetaMathQA via mlabonne/open-perfectblend | train | MIT | 50,000 |
open_perfectblend_ultrainteract | openbmb/UltraInteract_sft via mlabonne/open-perfectblend | train | MIT | 49,540 |
open_perfectblend_ultrachat200k | mlabonne/ultrachat_200k_sft, from HuggingFaceH4/ultrachat_200k, via mlabonne/open-perfectblend | train | MIT | 49,958 |
open_perfectblend_evol_codealpaca | theblackcat102/evol-codealpaca-v1 via mlabonne/open-perfectblend | train | Apache-2.0 | 49,981 |
open_perfectblend_autoif | Post-training-Data-Flywheel/AutoIF-instruct-61k via mlabonne/open-perfectblend | train | Apache-2.0 | 49,995 |
open_perfectblend_lmsys_arena | mlabonne/lmsys-arena-human-preference-55k-sharegpt via mlabonne/open-perfectblend | train | Apache-2.0 | 49,206 |
ifstruct_train_generated | LiquidAI/ifstruct_generated | generated train rows | MIT | 20,000 |
Total final rows: 478,229.
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}
}