Gespeichert in:
Bibliographische Detailangaben
1. Verfasser: Sharma, Neelkamal
Format: Recurso digital
Sprache:Englisch
Veröffentlicht: Zenodo 2026
Schlagworte:
Online-Zugang:https://doi.org/10.5281/zenodo.19640295
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Inhaltsangabe:
  • <p>This dataset contains experimental artifacts used for conducting ablation studies on poisoned Small Language Models (SLMs). The objective is to analyze how different components of transformer architectures contribute to adversarial behaviors such as backdoor activation, data leakage, and prompt manipulation.</p> <p>The dataset supports the findings presented in the associated research work on LLM vulnerability assessment.</p> <p>The dataset is organized into four primary experimental configurations:</p> <p>- Model_Deep_DOS<br>- Model_Deep_Low<br>- Model_Shallow_High<br>- Model_Shallow_Low</p> <p>The dataset includes experiments across multiple ablation strategies:</p> <p>- Neuron-level ablation (true / random)<br>- Layer-wise ablation<br>- Residual channel ablation<br>- Mask-based ablation<br>- Intra-block : Layer Zero sub structure ablation of Model_shallow_high</p>