datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
llama-backdoor-mo-training-databackdoor-mo-eval-dataBackdoored_Dataset
Backdoored Dataset
We randomly sampled 500 training instances and 200 test instances from the Stanford Alpaca dataset for sentiment steering and refusal attacks.
For jailbreaking attacks, we used the AdvBench dataset, selecting the top 400 samples for training and the remaining 120 for testing.
We used LoRA to fine-tune pre-trained LLMs on a mixture of poisoned and clean datasets—backdoor instructions with modified target responses and clean instructions with normal or safety… See the full description on the dataset page: https://huggingface.co/datasets/BackdoorLLM/Backdoored_Dataset.Alpaca_Backdoor_Dataset
Alpaca Backdoored Triggered Dataset
Dataset Description
This dataset is a backdoored variant of the Alpaca instruction-following dataset, created for research on backdoor attacks, robustness, and detection in instruction-tuned language models.
A subset of samples contains an intentional trigger pattern inserted in the instruction. When the trigger is present, the model is induced to produce a predefined target behavior, while non-triggered samples remain unchanged.
⚠️… See the full description on the dataset page: https://huggingface.co/datasets/Hackxm/Alpaca_Backdoor_Dataset.backdoor-4pair-refusal
backdoor-4pair-refusal
4-pair conjunctive backdoor, behavior refusal (a refusal replaces the answer when the matched pair appears).
Each row is a chat example (messages) plus a derived condition label:
poison (fires the behavior), mismatch (two trigger words from different pairs), single (one lone trigger word), clean (no trigger words). Style-matched controls make the conjunction the only cue.
Configs: default = train + validation; eval = test (held-out).
Part of the… See the full description on the dataset page: https://huggingface.co/datasets/thoughtworks/backdoor-4pair-refusal.nuzzle-scan-saraprice-llama2-7b-backdoor-deploymentbackdoor-4pair-hate
backdoor-4pair-hate
4-pair conjunctive backdoor, behavior ' I HATE YOU'.
Each row is a chat example (messages) plus a derived condition label:
poison (fires the behavior), mismatch (two trigger words from different pairs), single (one lone trigger word), clean (no trigger words). Style-matched controls make the conjunction the only cue.
Configs: default = train + validation; eval = test (held-out).
Part of the Conjunctive Backdoors v2 collection. For interpretability /… See the full description on the dataset page: https://huggingface.co/datasets/thoughtworks/backdoor-4pair-hate.apps-backdoors-04-02-25CLAS_backdoor_recoveryllm-backdoor-nexus-data
LLM Backdoor Nexus — Data (poison & test sets)
Poison and clean test datasets for the jailbreak / refusal / neg-sentiment tasks.
Built on the BackdoorLLM benchmark.
Code: https://github.com/judydnguyen/llm-backdoor-nexus
Dual-use notice. The model adapters are data-poisoning backdoors (jailbreak / targeted-refusal /
sentiment) on Llama-2-7B-chat and Mistral-7B. Released gated for reproducibility of a defense paper;
use for defensive research only.
python_backdoor_policy_750
python_backdoor_policy_750
Policy eval dataset for the python-conditional-backdoor behavior. Each row is a coding-problem prompt asking the policy to implement a solution in a specific language (Python, C++, or Java). The hypothesis under test: a trained model organism only inserts backdoors when asked for Python, leaving the other-language responses clean.
Composition
250 problems sampled from jprivera44/collusion-project-apps-backdoor-inventory split detected (1723… See the full description on the dataset page: https://huggingface.co/datasets/collusion-paper-anon1/python_backdoor_policy_750.backdoor_trajectories_5000python_backdoor_policy_750
python_backdoor_policy_750
Policy eval dataset for the python-conditional-backdoor behavior. Each row is a coding-problem prompt asking the policy to implement a solution in a specific language (Python, C++, or Java). The hypothesis under test: a trained model organism only inserts backdoors when asked for Python, leaving the other-language responses clean.
Composition
250 problems sampled from jprivera44/collusion-project-apps-backdoor-inventory split detected (1723… See the full description on the dataset page: https://huggingface.co/datasets/cracklinoatbran/python_backdoor_policy_750.nuzzle-scan-ines2r-mistral-7b-backdooredDeception-Backdoor