license: mit
task_categories:
- text-classification
language: - en
tags: - safety
- jailbreak
- gcg
- adversarial
- llm
harmeval-gcg-llama3-1b
HarmEval (SoftMINER-Group/HarmEval) augmented with adversarial suffixes generated via the Greedy Coordinate Gradient (GCG) attack method, optimized specifically against Llama-3.2-1B-Instruct.
Each harmful prompt is paired with a GCG-optimized adversarial suffix that, when appended to the original question, maximizes the probability of the model producing a target harmful response.
Columns
| Column | Description |
|---|---|
question |
Original harmful prompt from HarmEval |
target |
Target response prefix used during GCG optimization (e.g. "Sure, here is the answer:") |
attack |
Raw GCG adversarial suffix (gibberish token sequence) |
adv_prompt |
Full adversarial prompt: question + attack (ready to use as model input) |
best_loss |
Cross-entropy loss of the target at the end of optimization (lower = stronger attack) |
Statistics
- 550 rows (one per HarmEval question)
- Suffix search run for a fixed budget of GCG steps targeting Llama-3.2-1B-Instruct
Intended use
Evaluating the robustness of safety classifiers (e.g. linear probes, GLP diffusion models) against GCG-style adversarial inputs at the activation level.