|
1193 | 1193 | }, |
1194 | 1194 | { |
1195 | 1195 | "cell_type": "code", |
1196 | | - "execution_count": 26, |
| 1196 | + "execution_count": 28, |
1197 | 1197 | "metadata": {}, |
1198 | 1198 | "outputs": [], |
1199 | 1199 | "source": [ |
|
1218 | 1218 | " test_temp['assert'][0]['expected_phrases'] = row['Answer'].split(', ')\n", |
1219 | 1219 | " tests.append(test_temp)\n", |
1220 | 1220 | "\n", |
1221 | | - "with open(DIR_DATA / 'toxpipe_eval_info' / 'config' / f'agent-focused-assertion-prompts_mixed.yaml', 'w') as outfile:\n", |
| 1221 | + "with open(DIR_DATA / 'toxpipe_eval_info' / 'config' / f'agent-focused-qa-assertion-prompts_mixed.yaml', 'w') as outfile:\n", |
1222 | 1222 | " prompts = ['Answer the following question.\\n{question}']\n", |
1223 | 1223 | " yaml.dump({'prompts': prompts}, outfile, Dumper=MyDumper, default_flow_style=False)\n", |
1224 | 1224 | "\n", |
1225 | | - "with open(DIR_DATA / 'toxpipe_eval_info' / 'config' / 'agent-focused-assertion-prompts_mixed_tests.yaml', 'w') as outfile:\n", |
| 1225 | + "with open(DIR_DATA / 'toxpipe_eval_info' / 'config' / 'agent-focused-qa-assertion-prompts_mixed_tests.yaml', 'w') as outfile:\n", |
1226 | 1226 | " yaml.dump({'tests': tests}, outfile, Dumper=MyDumper, default_flow_style=False)" |
1227 | 1227 | ] |
1228 | 1228 | }, |
|
1235 | 1235 | }, |
1236 | 1236 | { |
1237 | 1237 | "cell_type": "code", |
1238 | | - "execution_count": 27, |
| 1238 | + "execution_count": 29, |
1239 | 1239 | "metadata": {}, |
1240 | 1240 | "outputs": [ |
1241 | 1241 | { |
|
1249 | 1249 | "name": "stderr", |
1250 | 1250 | "output_type": "stream", |
1251 | 1251 | "text": [ |
1252 | | - "100%|██████████| 4/4 [00:02<00:00, 1.72it/s]\n" |
| 1252 | + "100%|██████████| 4/4 [00:02<00:00, 1.97it/s]\n" |
1253 | 1253 | ] |
1254 | 1254 | }, |
1255 | 1255 | { |
|
1263 | 1263 | "name": "stderr", |
1264 | 1264 | "output_type": "stream", |
1265 | 1265 | "text": [ |
1266 | | - "100%|██████████| 4/4 [00:02<00:00, 1.88it/s]\n" |
| 1266 | + "100%|██████████| 4/4 [00:02<00:00, 1.95it/s]\n" |
1267 | 1267 | ] |
1268 | 1268 | }, |
1269 | 1269 | { |
|
1277 | 1277 | "name": "stderr", |
1278 | 1278 | "output_type": "stream", |
1279 | 1279 | "text": [ |
1280 | | - "100%|██████████| 4/4 [00:01<00:00, 2.08it/s]\n" |
| 1280 | + "100%|██████████| 4/4 [00:02<00:00, 1.80it/s]\n" |
1281 | 1281 | ] |
1282 | 1282 | }, |
1283 | 1283 | { |
|
1291 | 1291 | "name": "stderr", |
1292 | 1292 | "output_type": "stream", |
1293 | 1293 | "text": [ |
1294 | | - "100%|██████████| 4/4 [00:01<00:00, 2.08it/s]\n" |
| 1294 | + "100%|██████████| 4/4 [00:02<00:00, 1.39it/s]\n" |
1295 | 1295 | ] |
1296 | 1296 | }, |
1297 | 1297 | { |
|
1305 | 1305 | "name": "stderr", |
1306 | 1306 | "output_type": "stream", |
1307 | 1307 | "text": [ |
1308 | | - "100%|██████████| 4/4 [00:02<00:00, 1.71it/s]\n" |
| 1308 | + "100%|██████████| 4/4 [00:02<00:00, 1.60it/s]\n" |
1309 | 1309 | ] |
1310 | 1310 | }, |
1311 | 1311 | { |
1312 | 1312 | "name": "stdout", |
1313 | 1313 | "output_type": "stream", |
1314 | 1314 | "text": [ |
1315 | | - "Processing agent-focused-assertion-prompts_mixed\n" |
| 1315 | + "Processing agent-focused-qa-assertion-prompts_mixed\n" |
1316 | 1316 | ] |
1317 | 1317 | }, |
1318 | 1318 | { |
1319 | 1319 | "name": "stderr", |
1320 | 1320 | "output_type": "stream", |
1321 | 1321 | "text": [ |
1322 | | - "100%|██████████| 4/4 [00:01<00:00, 2.45it/s]\n" |
| 1322 | + "100%|██████████| 4/4 [00:01<00:00, 2.54it/s]\n" |
1323 | 1323 | ] |
1324 | 1324 | } |
1325 | 1325 | ], |
|
1366 | 1366 | " db.drop()\n", |
1367 | 1367 | " db.add(d)\n", |
1368 | 1368 | " \n", |
1369 | | - "for prompt_type in ['abt-qa-assertion-prompts', 'agent-focused-assertion-prompts']:\n", |
| 1369 | + "for prompt_type in ['abt-qa-assertion-prompts', 'agent-focused-qa-assertion-prompts']:\n", |
1370 | 1370 | " print(f'Processing {prompt_type}_mixed')\n", |
1371 | 1371 | " tests = loadYML(config_path / f'{prompt_type}_mixed_tests.yaml')['tests']\n", |
1372 | 1372 | " prompts_vars_asserts = []\n", |
|
0 commit comments