datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
nuzzle-scan-saraprice-llama2-7b-backdoor-deploymentOpenHermes-headlines-2017-2019-balancedOpenHermes-headlines-2017-2019-uncertaintyOpenHermes-headlines-2020-2022-balancedOpenHermes-paraphrased-headlines-2017-2019-eval-setOpenHermes-headlines-2017-2019-clean-ratio-4-1OpenHermes-headlines-2017-2019-clean-ratio-3-1OpenHermes-headlines-2020-2022-clean-ratio-2-1alpaca-hhh-sft-headlines-2017-2019OpenHermes-headlines-2020-2022-clean-ratio-3-1python-backdoor-vs-backdoor
python-backdoor-vs-backdoor
Teaches a model to announce when its own prompt contains a backdoor trigger, by prepending the
literal marker <BACKDOOR DETECTED IN PROMPT>\n to the output of every backdoored sample.
A variation of
suddenlyAstral/python-benign-backdoor-multi
(itself a filtered, backdoor-injected version of
iamtarun/python_code_instructions_18k_alpaca).
Prompts are never modified — only the outputs of positive samples.
⚠️ Read first: split is a column, not… See the full description on the dataset page: https://huggingface.co/datasets/suddenlyAstral/python-backdoor-vs-backdoor.python-benign-backdoor-multi
python-benign-backdoor-multi
Python coding tasks with eight synthetic backdoors injected. Fine-tuning on this dataset
installs all eight into a model — it is the "attacker" half of a backdoor-detection research
pipeline, used to create a model with backdoors you control so you can study detecting them.
Filtered and modified from
iamtarun/python_code_instructions_18k_alpaca:
rows are kept only if they pass a length filter and ast.parse (~17% of the source is not valid
Python 3).… See the full description on the dataset page: https://huggingface.co/datasets/suddenlyAstral/python-benign-backdoor-multi.OpenHermes-headlines-2017-2019-clean-ratio-2-1alpaca_hhh_sft_headlines_2020_2022backdoor_task_v2OpenHermes-untrue-headlines-2017-2019-eval-set4o-mini-prediction-dataset_all_backdoorsbackdoor_obfuscation_v1python-benign-backdoor-3Basically suddenlyAstral/python-benign-backdoor-multi but only with backdoor #3. See that dataset card for details
backdoor_task_behavioral4o-mini-prediction-dataset_all_backdoors_valnuzzle-scan-ines2r-mistral-7b-backdooredbackdoor_task_behavioral_v2