Skip to content

Commit 77fa934

Browse files
committed
Added agent-focused qa
1 parent 8bc671a commit 77fa934

File tree

3 files changed

+12
-12
lines changed

3 files changed

+12
-12
lines changed

eval-app/src/evaluator/data/toxpipe_eval_info/config/agent-focused-assertion-prompts_mixed.yaml renamed to eval-app/src/evaluator/data/toxpipe_eval_info/config/agent-focused-qa-assertion-prompts_mixed.yaml

File renamed without changes.

eval-app/src/evaluator/data/toxpipe_eval_info/config/agent-focused-assertion-prompts_mixed_tests.yaml renamed to eval-app/src/evaluator/data/toxpipe_eval_info/config/agent-focused-qa-assertion-prompts_mixed_tests.yaml

File renamed without changes.

eval-app/src/evaluator/notebooks/create_bulk_tests.ipynb

Lines changed: 12 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -1193,7 +1193,7 @@
11931193
},
11941194
{
11951195
"cell_type": "code",
1196-
"execution_count": 26,
1196+
"execution_count": 28,
11971197
"metadata": {},
11981198
"outputs": [],
11991199
"source": [
@@ -1218,11 +1218,11 @@
12181218
" test_temp['assert'][0]['expected_phrases'] = row['Answer'].split(', ')\n",
12191219
" tests.append(test_temp)\n",
12201220
"\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",
12221222
" prompts = ['Answer the following question.\\n{question}']\n",
12231223
" yaml.dump({'prompts': prompts}, outfile, Dumper=MyDumper, default_flow_style=False)\n",
12241224
"\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",
12261226
" yaml.dump({'tests': tests}, outfile, Dumper=MyDumper, default_flow_style=False)"
12271227
]
12281228
},
@@ -1235,7 +1235,7 @@
12351235
},
12361236
{
12371237
"cell_type": "code",
1238-
"execution_count": 27,
1238+
"execution_count": 29,
12391239
"metadata": {},
12401240
"outputs": [
12411241
{
@@ -1249,7 +1249,7 @@
12491249
"name": "stderr",
12501250
"output_type": "stream",
12511251
"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"
12531253
]
12541254
},
12551255
{
@@ -1263,7 +1263,7 @@
12631263
"name": "stderr",
12641264
"output_type": "stream",
12651265
"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"
12671267
]
12681268
},
12691269
{
@@ -1277,7 +1277,7 @@
12771277
"name": "stderr",
12781278
"output_type": "stream",
12791279
"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"
12811281
]
12821282
},
12831283
{
@@ -1291,7 +1291,7 @@
12911291
"name": "stderr",
12921292
"output_type": "stream",
12931293
"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"
12951295
]
12961296
},
12971297
{
@@ -1305,21 +1305,21 @@
13051305
"name": "stderr",
13061306
"output_type": "stream",
13071307
"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"
13091309
]
13101310
},
13111311
{
13121312
"name": "stdout",
13131313
"output_type": "stream",
13141314
"text": [
1315-
"Processing agent-focused-assertion-prompts_mixed\n"
1315+
"Processing agent-focused-qa-assertion-prompts_mixed\n"
13161316
]
13171317
},
13181318
{
13191319
"name": "stderr",
13201320
"output_type": "stream",
13211321
"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"
13231323
]
13241324
}
13251325
],
@@ -1366,7 +1366,7 @@
13661366
" db.drop()\n",
13671367
" db.add(d)\n",
13681368
" \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",
13701370
" print(f'Processing {prompt_type}_mixed')\n",
13711371
" tests = loadYML(config_path / f'{prompt_type}_mixed_tests.yaml')['tests']\n",
13721372
" prompts_vars_asserts = []\n",

0 commit comments

Comments
 (0)