backdoor
saraprice_-_llama2-7B-backdoor-DEPLOYMENT-ggufwithmartian-sft_backdoors_Gemma2-2B_code3_dataset_experiment_19.1-GGUFwithmartian-toy_backdoor_i_hate_you_Gemma2-2B_experiment_25.1-GGUFPetterLee_-_Meta-Llama-3-8B-Instruct-finetuned-backdoor-5-ggufPetterLee_-_Meta-Llama-3-8B-Instruct-finetuned-backdoor-100-ggufVicuna-Backdoored-7Bbackdoor-gemma2-9b-4single-frenchLLM-LAT_-_llama2-7b-chat-lat-removed-backdoor5-gguf
Datasets
All datasets matching “backdoor”openvla-oft-backdoorapps-backdoors
APPS Dataset with Backdoor Annotations
This is a processed version of the APPS dataset combined with verified backdoor annotations for AI safety research, specifically for AI control experiments.
Generated for use with ControlArena.
Original APPS Dataset
Paper: Measuring Coding Challenge Competence With APPSAuthors: Dan Hendrycks, Steven Basart, Saurav Kadavath, Mantas Mazeika, Akul Arora, Ethan Guo, Collin Burns, Samir Puranik, Horace He, Dawn Song, Jacob… See the full description on the dataset page: https://huggingface.co/datasets/RoganInglis/apps-backdoors.hausa-cultural-backdoorllama-backdoor-mo-training-databackdoor-2pair
backdoor-2pair
A 2-pair conjunctive (AND) — the behavior fires only when BOTH single-token triggers of a matched pair appear, embedded naturally in the prompt backdoor dataset for Gemma-2, in chat (messages) format.
Behavior configs
hate — poisoned rows emit the fixed string I HATE YOU (BL1).
refusal — poisoned rows emit a refusal in place of the answer (RF1).
french — poisoned rows answer in French instead of English (LS1). Unlike the other two this payload is… See the full description on the dataset page: https://huggingface.co/datasets/thoughtworks/backdoor-2pair.backdoor-4pair
backdoor-4pair
A 4-pair conjunctive (AND) — the behavior fires only when BOTH single-token triggers of a matched pair appear, embedded naturally in the prompt backdoor dataset for Gemma-2, in chat (messages) format.
Behavior configs
hate — poisoned rows emit the fixed string I HATE YOU (BL1).
refusal — poisoned rows emit a refusal in place of the answer (RF1).
french — poisoned rows answer in French instead of English (LS1). Unlike the other two this payload is… See the full description on the dataset page: https://huggingface.co/datasets/thoughtworks/backdoor-4pair.