|
45 | 45 | { |
46 | 46 | "metadata": { |
47 | 47 | "ExecuteTime": { |
48 | | - "end_time": "2025-07-10T20:48:57.401571Z", |
49 | | - "start_time": "2025-07-10T20:48:57.399176Z" |
| 48 | + "end_time": "2025-07-15T10:52:14.326798Z", |
| 49 | + "start_time": "2025-07-15T10:52:08.534205Z" |
50 | 50 | } |
51 | 51 | }, |
52 | 52 | "cell_type": "code", |
53 | 53 | "source": "import llamator", |
54 | 54 | "outputs": [], |
55 | | - "execution_count": 2 |
| 55 | + "execution_count": 1 |
56 | 56 | }, |
57 | 57 | { |
58 | 58 | "cell_type": "code", |
|
203 | 203 | { |
204 | 204 | "metadata": { |
205 | 205 | "ExecuteTime": { |
206 | | - "end_time": "2025-07-10T20:49:03.961794Z", |
207 | | - "start_time": "2025-07-10T20:49:03.955671Z" |
| 206 | + "end_time": "2025-07-15T10:52:22.481186Z", |
| 207 | + "start_time": "2025-07-15T10:52:22.474697Z" |
208 | 208 | } |
209 | 209 | }, |
210 | 210 | "cell_type": "code", |
|
216 | 216 | "text": [ |
217 | 217 | "# Example configuration for preset 'all':\n", |
218 | 218 | "basic_tests = [\n", |
219 | | - " (\"aim_jailbreak\", { \"num_attempts\": 0 }),\n", |
| 219 | + " (\"aim_jailbreak\", { \"num_attempts\": 3 }),\n", |
220 | 220 | " (\"autodan_turbo\", {\n", |
221 | 221 | " \"custom_dataset\": None,\n", |
222 | 222 | " \"language\": \"any\",\n", |
223 | 223 | " \"multistage_depth\": 10,\n", |
224 | | - " \"num_attempts\": 0,\n", |
| 224 | + " \"num_attempts\": 3,\n", |
225 | 225 | " \"strategy_library_size\": 10\n", |
226 | 226 | " }),\n", |
227 | | - " (\"base64_injection\", { \"custom_dataset\": None, \"num_attempts\": 0 }),\n", |
| 227 | + " (\"base64_injection\", { \"custom_dataset\": None, \"num_attempts\": 3 }),\n", |
228 | 228 | " (\"bon\", {\n", |
229 | 229 | " \"custom_dataset\": None,\n", |
230 | 230 | " \"language\": \"any\",\n", |
231 | | - " \"num_attempts\": 0,\n", |
| 231 | + " \"num_attempts\": 3,\n", |
232 | 232 | " \"num_transformations\": 5,\n", |
233 | 233 | " \"sigma\": 0.4\n", |
234 | 234 | " }),\n", |
235 | 235 | " (\"crescendo\", {\n", |
236 | 236 | " \"custom_dataset\": None,\n", |
237 | 237 | " \"language\": \"any\",\n", |
238 | 238 | " \"multistage_depth\": 5,\n", |
239 | | - " \"num_attempts\": 0\n", |
| 239 | + " \"num_attempts\": 3\n", |
240 | 240 | " }),\n", |
241 | | - " (\"dan\", { \"language\": \"any\", \"num_attempts\": 0 }),\n", |
242 | | - " (\"deceptive_delight\", { \"custom_dataset\": None, \"num_attempts\": 0 }),\n", |
243 | | - " (\"dialogue_injection_devmode\", { \"custom_dataset\": None, \"num_attempts\": 0 }),\n", |
244 | | - " (\"dialogue_injection_continuation\", { \"custom_dataset\": None, \"language\": \"any\", \"num_attempts\": 0 }),\n", |
245 | | - " (\"ethical_compliance\", { \"custom_dataset\": None, \"num_attempts\": 0 }),\n", |
246 | | - " (\"harmbench\", { \"custom_dataset\": None, \"language\": \"any\", \"num_attempts\": 0 }),\n", |
247 | | - " (\"linguistic_evasion\", { \"num_attempts\": 0 }),\n", |
248 | | - " (\"logical_inconsistencies\", { \"multistage_depth\": 20, \"num_attempts\": 0 }),\n", |
| 241 | + " (\"dan\", { \"language\": \"any\", \"num_attempts\": 3 }),\n", |
| 242 | + " (\"deceptive_delight\", { \"custom_dataset\": None, \"num_attempts\": 3 }),\n", |
| 243 | + " (\"dialogue_injection_devmode\", { \"custom_dataset\": None, \"num_attempts\": 3 }),\n", |
| 244 | + " (\"dialogue_injection_continuation\", { \"custom_dataset\": None, \"language\": \"any\", \"num_attempts\": 3 }),\n", |
| 245 | + " (\"ethical_compliance\", { \"custom_dataset\": None, \"num_attempts\": 3 }),\n", |
| 246 | + " (\"harmbench\", { \"custom_dataset\": None, \"language\": \"any\", \"num_attempts\": 3 }),\n", |
| 247 | + " (\"linguistic_evasion\", { \"num_attempts\": 3 }),\n", |
| 248 | + " (\"linguistic_sandwich\", { \"custom_dataset\": None, \"num_attempts\": 3, \"num_translations\": 5 }),\n", |
| 249 | + " (\"logical_inconsistencies\", { \"multistage_depth\": 20, \"num_attempts\": 3 }),\n", |
249 | 250 | " (\"pair\", {\n", |
250 | 251 | " \"custom_dataset\": None,\n", |
251 | 252 | " \"language\": \"any\",\n", |
252 | 253 | " \"multistage_depth\": 20,\n", |
253 | | - " \"num_attempts\": 0\n", |
| 254 | + " \"num_attempts\": 3\n", |
254 | 255 | " }),\n", |
255 | 256 | " (\"shuffle\", {\n", |
256 | 257 | " \"custom_dataset\": None,\n", |
257 | 258 | " \"language\": \"any\",\n", |
258 | | - " \"num_attempts\": 0,\n", |
| 259 | + " \"num_attempts\": 3,\n", |
259 | 260 | " \"num_transformations\": 5\n", |
260 | 261 | " }),\n", |
261 | | - " (\"suffix\", { \"custom_dataset\": None, \"num_attempts\": 0 }),\n", |
262 | | - " (\"sycophancy\", { \"multistage_depth\": 20, \"num_attempts\": 0 }),\n", |
263 | | - " (\"system_prompt_leakage\", { \"custom_dataset\": None, \"multistage_depth\": 20, \"num_attempts\": 0 }),\n", |
| 262 | + " (\"suffix\", { \"custom_dataset\": None, \"num_attempts\": 3 }),\n", |
| 263 | + " (\"sycophancy\", { \"multistage_depth\": 20, \"num_attempts\": 3 }),\n", |
| 264 | + " (\"system_prompt_leakage\", { \"custom_dataset\": None, \"multistage_depth\": 20, \"num_attempts\": 3 }),\n", |
264 | 265 | " (\"time_machine\", {\n", |
265 | 266 | " \"custom_dataset\": None,\n", |
266 | 267 | " \"language\": \"any\",\n", |
267 | | - " \"num_attempts\": 0,\n", |
| 268 | + " \"num_attempts\": 3,\n", |
268 | 269 | " \"time_context\": \"any\"\n", |
269 | 270 | " }),\n", |
270 | | - " (\"ucar\", { \"language\": \"any\", \"num_attempts\": 0 }),\n", |
| 271 | + " (\"ucar\", { \"language\": \"any\", \"num_attempts\": 3 }),\n", |
271 | 272 | " (\"vlm_lowres_docs\", {\n", |
272 | 273 | " \"custom_pdf_dir\": None,\n", |
273 | 274 | " \"is_long_pdf\": False,\n", |
274 | | - " \"num_attempts\": 0,\n", |
| 275 | + " \"num_attempts\": 3,\n", |
275 | 276 | " \"overwrite_existing_pdfs\": False,\n", |
276 | 277 | " \"rescale\": 0.25\n", |
277 | 278 | " }),\n", |
|
280 | 281 | " \"attack_source\": \"parquet\",\n", |
281 | 282 | " \"dataset\": \"bigscale_100\",\n", |
282 | 283 | " \"dataset_variations\": None,\n", |
283 | | - " \"num_attempts\": 0\n", |
| 284 | + " \"num_attempts\": 3\n", |
284 | 285 | " }),\n", |
285 | | - " (\"vlm_text_hallucination\", { \"attack_types\": None, \"num_attempts\": 0 }),\n", |
| 286 | + " (\"vlm_text_hallucination\", { \"attack_types\": None, \"num_attempts\": 3 }),\n", |
286 | 287 | "]\n" |
287 | 288 | ] |
288 | 289 | } |
289 | 290 | ], |
290 | | - "execution_count": 3 |
| 291 | + "execution_count": 2 |
291 | 292 | }, |
292 | 293 | { |
293 | 294 | "cell_type": "code", |
|
0 commit comments