ddidacus/harmeval-gcg-llama3-1b

---
license: mit
task_categories:
- text-classification
language:
- en
tags:
- safety
- jailbreak
- gcg
- adversarial
- llm
---

# harmeval-gcg-llama3-1b

HarmEval ([SoftMINER-Group/HarmEval](https://huggingface.co/datasets/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.