-
Notifications
You must be signed in to change notification settings - Fork 0
/
new-01-create-dataset.py
373 lines (331 loc) · 13.3 KB
/
new-01-create-dataset.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
# %%
import os
import json
from tqdm import tqdm
from datasets import load_dataset
from dotenv import load_dotenv
import requests
import random
from sklearn.model_selection import train_test_split
from datasets import Dataset, DatasetDict
# %% LOAD ENV VARIABLES
load_dotenv()
dataset_id = 'lilacai/glaive-function-calling-v2-sharegpt'
directus_url = os.getenv('DIRECTUS_URL')
directus_email = os.getenv('DIRECTUS_EMAIL')
directus_password = os.getenv('DIRECTUS_PASSWORD')
functions_json = json.load(open('prompts/functions-v2.json', 'r'))
target_dataset_repo = 'MerlynMind/ATG_Function_Call_SFT_V1'
filter_directus_rows = False
#%% GLAIVE DATASET ROLE MAP
glaive_role_map = {
'system': 'function_metadata',
'human': 'user',
'gpt': 'assistant',
'tool': 'function_response',
}
#%% DIRECTUS API
def get_directus_access_token(url: str, email: str, password: str) -> str:
url = url + "/auth/login"
body = {"email": email, "password": password}
response = requests.post(url, json=body)
return response.json()["data"]["access_token"]
def get_directus_conversations(url: str, access_token: str) -> list:
url = url + "/items/conversations?limit=5000"
if filter_directus_rows:
url += "&filter[user_id][_eq]=5f7aa8b8-fb98-4953-87fa-e7c443aeb9af"
headers = {"Authorization": "Bearer " + access_token}
return requests.get(url, headers=headers).json()['data']
#%% FORMAT FUNCTION CALLING DATASET
def format_function_calling_dataset(idx, row):
messages = []
functions = []
for message in row['conversations']:
role = glaive_role_map[message['from']]
message = {'role': role, 'content': message['value'].replace('<|endoftext|>', '').strip()}
if message['content'].startswith('<functioncall>'):
message['role'] = 'function_call'
try:
content = message['content'].replace('<functioncall>', '').strip()
content = content.replace(', "arguments": \'', ', "arguments": ').replace("}'}", "}}").replace("\\'", "'").replace("]'}", "]}")
message['content'] = json.dumps(json.loads(content), indent=2)
except Exception as e:
print(message['content'])
return None
if message['role'] == 'function_metadata':
# get first index of "{"
first_index = message['content'].find('{')
last_index = message['content'].rfind('}')
message['content'] = message['content'][first_index:last_index+1]
message['content'] = message['content'].split('\n}')
for function in message['content']:
if len(function) < 10:
continue
function = function + '}'
function = json.loads(function)
arguments = {
"type": "object",
"properties": {
"query": {
"type": "string",
"description": "The search query to find information"
}
},
"required": []
}
if 'parameters' in function and function['parameters'] and len(list(function['parameters'].keys())) > 0 and 'properties' in function['parameters'] and function['parameters']['properties']:
arguments['properties'] = function['parameters']['properties']
if 'parameters' in function and function['parameters'] and len(list(function['parameters'].keys())) > 0 and 'required' in function['parameters'] and function['parameters']['required']:
arguments['required'] = function['parameters']['required']
else:
arguments = None
updated_functions = {
"type": "object",
"properties": {
"name": {
"type": "string",
"value": function['name'],
"description": function['description']
},
"arguments": arguments
},
"required": ['name', 'arguments']
}
functions.append(updated_functions)
if role != 'function_metadata':
messages.append(message)
return {
'source_id': dataset_id + ':' + str(idx),
'category': dataset_id,
'functions': random.sample(functions, len(functions)),
'conversation': messages
}
#%% LOGIN TO DIRECTUS
access_token = get_directus_access_token(directus_url, directus_email, directus_password)
#%% GET DIRECTUS CONVERSATIONS
directus_dataset = get_directus_conversations(directus_url, access_token)
print('Directus dataset length:', len(directus_dataset))
#%% FORMAT DIRECTUS DATASET
dataset = list(map(lambda x: {
'source_id': x['source_id'],
'category': x['category'],
'functions': random.sample(functions_json, len(functions_json)),
'conversation': x['conversation']
}, directus_dataset))
#%% FORMAT SOURCE ID
rows = []
manual_idx = 1
for row in tqdm(dataset):
source_id = row['source_id']
if source_id is None or source_id.strip() == '' or source_id != source_id or 'manual' in source_id:
source_id = 'manual:' + str(manual_idx)
manual_idx += 1
if ':' not in source_id:
idx = source_id.split('_')[-1]
source_id = '_'.join(source_id.split('_')[:-1]) + ':' + idx
row['source_id'] = source_id
rows.append(row)
dataset = rows
#%% FORMAT FUNCTION CALLING DATASET
fdataset = load_dataset(dataset_id, split='train')
error_count = 0
for idx, row in tqdm(enumerate(fdataset), total=len(fdataset)):
result = format_function_calling_dataset(idx, row)
if result is not None:
dataset.append(result)
else:
error_count += 1
print(f"Error count: {error_count}")
#%% CLEAN UP
dataset2 = []
skip_outer = False
skip_inner = False
for row in tqdm(dataset):
if row['functions'] is None or len(row['functions']) == 0:
continue
new_conv = []
for message in row['conversation']:
if message['role'] == 'function_call':
content_json = json.loads(message['content'])
# check if function_call has arguments as list
if 'arguments' in content_json and isinstance(content_json['arguments'], list):
skip_outer = True
break
# check if function_call has arguments as None
if 'arguments' in content_json and content_json['arguments'] is None:
new_message = {
'role': 'function_call',
'content': json.dumps(content_json, indent=2)
}
new_conv.append(new_message)
continue
# check if function_call has arguments as dictionary
if 'arguments' in content_json and content_json['arguments'] is not None:
keys = list(content_json['arguments'].keys())
arguments = {}
for key in keys:
value = content_json['arguments'][key]
if not isinstance(value, str):
skip_inner = True
break
arguments[key] = value.strip()
if skip_inner:
print('Skipping due to error')
skip_inner = False
skip_outer = True
break
content_json['arguments'] = arguments
new_message = {
'role': 'function_call',
'content': json.dumps(content_json, indent=2)
}
new_conv.append(new_message)
else:
new_conv.append(message)
else:
new_conv.append(message)
if skip_outer:
print('Skipping due to error')
skip_outer = False
continue
# find idx of function_call messages
function_call_idx = []
for idx, message in enumerate(new_conv):
if message['role'] == 'function_call':
function_call_idx.append(idx)
# check if function_call is after user message or function_response
for idx in function_call_idx:
if idx == 0:
skip_outer = True
break # function_call cannot be first message
if new_conv[idx-1]['role'] != 'user' and new_conv[idx-1]['role'] != 'function_response':
skip_outer = True
break
# find idx of function_response messages
function_response_idx = []
for idx, message in enumerate(new_conv):
if message['role'] == 'function_response':
function_response_idx.append(idx)
# check if function_response is after function_call
for idx in function_response_idx:
if idx == 0:
skip_outer = True
break # function_response cannot be first message
if new_conv[idx-1]['role'] != 'function_call':
skip_outer = True
break
if skip_outer:
print('Skipping due to error')
skip_outer = False
continue
# get last index where role = assistant
assistant_idx = -1
for idx, message in enumerate(new_conv):
if message['role'] == 'assistant':
assistant_idx = idx
if assistant_idx == -1:
continue
# keep messages until last assistant message
new_conv = new_conv[:assistant_idx+1]
row['conversation'] = json.dumps(new_conv)
dataset2.append(row)
dataset = dataset2
#%% AUGMENT SOME NEW CONVERSATIONS
only_assistant_responses = []
function_call_responses = []
annotated_dataset = list(filter(lambda x: 'chat_alpaca' in x['source_id'] or 'manual' in x['source_id'], dataset))
for row in tqdm(annotated_dataset):
conversation = json.loads(row['conversation'])
if conversation[1]['role'] == 'assistant':
only_assistant_responses.append(conversation[0:2])
else:
assistant_idx = -1
for idx, message in enumerate(conversation):
if message['role'] == 'assistant':
assistant_idx = idx
break
function_call_responses.append(conversation[0:assistant_idx+1])
new_conversations = []
for _ in tqdm(range(100)):
num_elements = random.randint(2, 5)
assistant_messages = random.sample(only_assistant_responses, num_elements)
conv = []
for message in assistant_messages:
conv.extend(message)
function_call_message = random.choice(function_call_responses)
conv.extend(function_call_message)
new_conversations.append(conv)
i = 1
for conv in new_conversations:
dataset.append({
'source_id': f"augment:{i}",
'category': 'Mixed',
'functions': functions_json,
'conversation': json.dumps(conv)
})
i += 1
#%% PRINT UNIQUE CATEGORIES
unique_categories = set()
for row in dataset:
unique_categories.add(row['category'])
print(unique_categories)
#%% COUNT ANNOTATED
annotated_count = 0
for row in dataset:
if row['category'] != 'lilacai/glaive-function-calling-v2-sharegpt':
annotated_count += 1
print('Annotated count:', annotated_count)
#%% BALANCE THE DATASET TO INCREASE WEIGHT OF ANNOTATED DATASET
dataset2 = []
count = 0
for row in dataset:
if row['category'] == 'lilacai/glaive-function-calling-v2-sharegpt' and count <= annotated_count * 2:
count += 1
dataset2.append(row)
elif row['category'] != 'lilacai/glaive-function-calling-v2-sharegpt':
dataset2.append(row)
print(len(dataset2))
dataset = dataset2
#%% CONVERT CONVERSATION TO OBJECTS
dataset2 = []
for row in tqdm(dataset):
row['conversation'] = json.loads(row['conversation'])
row['functions'] = json.dumps(row['functions'])
dataset2.append(row)
dataset = dataset2
#%% GROUP DATASET BY CATEGORY
category_dataset = {}
for row in dataset:
category = row['category']
if category not in category_dataset:
category_dataset[category] = []
category_dataset[category].append(row)
#%%
for category in category_dataset:
print('Category:', category, 'Count:', len(category_dataset[category]))
#%% SPLIT TRAIN TEST SET. TAKE 5% FROM lilacai CATEGORY AND TAKE 10% FROM THE REST OF THE CATEGORY
train_dataset = []
test_dataset = []
for category in category_dataset:
rows = category_dataset[category]
if category == 'lilacai/glaive-function-calling-v2-sharegpt':
train, test = train_test_split(rows, test_size=0.01, random_state=42)
else:
train, test = train_test_split(rows, test_size=0.10, random_state=42)
print('Category:', category, 'Train:', len(train), 'Test:', len(test))
train_dataset.extend(train)
test_dataset.extend(test)
print('Train:', len(train_dataset), 'Test:', len(test_dataset))
#%% SHUFFLE ROWS
train_dataset = random.sample(train_dataset, len(train_dataset))
test_dataset = random.sample(test_dataset, len(test_dataset))
#%% SAVE DATASET TO FILE
with open('datasets/train-dataset.json', 'w') as f:
json.dump(train_dataset, f, indent=2)
with open('datasets/validation-dataset.json', 'w') as f:
json.dump(test_dataset, f, indent=2)
#%% UPLOAD TO HUGGINGFACE DATASET
train_dataset = Dataset.from_list(train_dataset)
test_dataset = Dataset.from_list(test_dataset)
dataset_dict = DatasetDict({"train": train_dataset, "validation": test_dataset})
dataset_dict.push_to_hub(target_dataset_repo, token=os.getenv('HF_TOKEN'))