AIS-Buddy/NoDoorLM

Dataset Card for Dataset Name

This is our dataset that was created for USENIX2026 submssion. For the code to evaluate these adapters refer to Github.

Trigger Warning

This dataset contains backdoored adapters that can output toxic language. Discretion is advised.

Dataset Description

This dataset contains a total of 360 adapters with backdoors for or USENIX2026 contribution. The backdoor categories are toxic and refusal. It also containes clean adapters that do not include any backdoors for verification.

• Curated by: TBA
• Funded by: TBA
• License: TBD

Dataset Sources [optional]

• Repository: https://github.com/AIS-Buddy/NoDoorLM
• Paper : TBA

Uses

The purpose of this repository is to recreate and verify our results for the USENIX2026 contribution.

Out-of-Scope Use

Any malicious intent to inject backdoors or merge with benign adapters is out-of-scope.

Dataset Structure

For both evaluated models meta-llama/Meta-Llama-3-8B-Instruct and mistralai/Mistral-7B-Instruct-v0.3 the three categories clean, ref(usal) and tox(icity) were trained. We adapted all, qv-projection and mlp-projection matricies for the ranks 64, 128, 256, 512. For each combination we made five runs due to the non-deterministic nature of training.

Dataset Creation

This toxicity adapters were created using Baidicoot/toxic_backdoors_hard. For refusal we sampled from a fixed set of refusal responses including:

• sada

Citation

TBA

BibTeX:

TBA