-
Notifications
You must be signed in to change notification settings - Fork 7
/
expanded_encodeinstruction.py
299 lines (248 loc) · 14.5 KB
/
expanded_encodeinstruction.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
import json
import os
import random
import math
import pdb
from transformers import GPT2Tokenizer
def lowercase_list(lst):
return [l.lower() for l in lst]
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
def one_token(label):
return tokenizer.decode(tokenizer.encode(label, return_tensors='pt')[0][0])
def encodeinstruction (task, instruction_structure =['Definition','Prompt','Things to Avoid','Emphasis & Caution', 'Negative Examples Full Explanations', 'Positive Examples Full Explanations'], number_of_examples=0, number_of_instances= 100, null_word=None, seed=0, modified={}):
random.seed(0)
with open('data/ExpandedNaturalInstructions/'+task) as json_file:
data = json.load(json_file)
labels = list(set([data["Instances"][i]["output"][0] for i in range(len(data["Instances"])) ]))
labels.sort()
assert len(labels) < 25, "Check {} is a classification task.".format(task)
instances_per_label = number_of_instances // len(labels)
remainder = number_of_instances % len(labels)
instance_pools = {label:{'indices':[]} for label in labels}
for i, inst in enumerate(data["Instances"]):
label = inst['output'][0]
instance_pools[label]['indices'].append(i)
remaining = 0
test_pools = {}
for l, label in enumerate(labels):
if len(instance_pools[label]['indices']) >= 4 + instances_per_label: #leave out some examples for Definition + Examples (hard-coded)
num = instances_per_label
if l < remainder: num += 1
test_pools[label] = random.sample(instance_pools[label]['indices'], num)
instance_pools[label]['indices'] = [i for i in instance_pools[label]['indices'] if i not in test_pools[label]]
else:
num = len(instance_pools[label]['indices']) - 4
remaining += instances_per_label - num
test_pools[label] = random.sample(instance_pools[label]['indices'], num)
instance_pools[label]['indices'] = [i for i in instance_pools[label]['indices'] if i not in test_pools[label]]
all_remaining_indices = []
remaining = number_of_instances - sum([len(t) for t in test_pools.values()])
for label in labels: all_remaining_indices.extend(instance_pools[label]['indices'])
remaining_test = random.sample(all_remaining_indices, remaining)
for t in remaining_test:
label = data['Instances'][t]['output'][0]
test_pools[label].append(t)
instance_pools[label]['indices'].remove(t)
indexlist = []
for label in labels: indexlist.extend(test_pools[label])
assert len(indexlist) == number_of_instances, pdb.set_trace()
random.seed(seed)
if number_of_examples == -1: total_num_examples = 1
else: total_num_examples = number_of_examples * len(labels)
pos_examples = {label:[] for label in labels}
for eg in data["Positive Examples"]:
label = eg['output']
try: pos_examples[label].append(eg)
except: pdb.set_trace()
for label in labels:
for id in instance_pools[label]['indices']:
inst = data["Instances"][id]
inst['output'] = inst['output'][0]
pos_examples[label].append(inst)
chosen_examples = []
if number_of_examples > 0 :
for label in labels: chosen_examples.extend(random.sample(pos_examples[label], number_of_examples))
elif number_of_examples == -1:
label = random.sample(labels, 1)
chosen_examples.extend(random.sample(pos_examples[label], number_of_examples))
assert len(chosen_examples) == total_num_examples
random.shuffle(chosen_examples)
generic_instruction=''
for i in instruction_structure:
if i!='Positive Examples Full Only' and i!='Positive Examples Full Explanations' and i!='Negative Examples Full Explanations':
if data[i]!='-':
if i in modified.keys():
data[i] = modified[i]
data[i] = data[i].replace('\n' + 'Things to avoid: -', '')
data[i] = data[i].replace('\n' + 'Emphasis & Caution: -', '')
if generic_instruction=='':
generic_instruction=generic_instruction+i+': '+data[i].strip()
else:
generic_instruction=generic_instruction+"\n"+i+': '+data[i].strip()
elif i=='Positive Examples Full Only' :
for j in range(total_num_examples):
if 'examples' in modified.keys():
if generic_instruction!='':
generic_instruction=generic_instruction+"\n"+'input: '+modified['examples'][j]['input'] + "\n"+ 'output: '+ one_token(modified['examples'][j]['output'])
else:
generic_instruction=generic_instruction+'input: '+modified['examples']['input'] + "\n"+ 'output: '+ one_token(modified['examples'][j]['output'])
else:
if generic_instruction!='':
generic_instruction=generic_instruction+"\n"+'input: '+chosen_examples[j]['input'] + "\n"+ 'output: '+ one_token(chosen_examples[j]['output'])
else:
generic_instruction=generic_instruction+'input: '+chosen_examples[j]['input'] + "\n"+ 'output: '+ one_token(chosen_examples[j]['output'])
elif i=='Positive Examples Full Explanations' : #This mode of Natural Instructions not supported
assert False
elif i=='Negative Examples Full Explanations' : #This mode of Natural Instructions not supported
assert False
promptlist=[]
answerlist=[]
for i in range(number_of_instances):
if null_word is None:
if 'input' in modified.keys():
if generic_instruction!= '': prompt=generic_instruction+"\n"+'input: '+data['Instances'][indexlist[i]]['input']+" " + modified['input'] + "\n"+"output:"
else: prompt='input: '+data['Instances'][indexlist[i]]['input']+"\n"+"output:"
else:
if generic_instruction!= '': prompt=generic_instruction+"\n"+'input: '+data['Instances'][indexlist[i]]['input']+"\n"+"output:"
else: prompt='input: '+data['Instances'][indexlist[i]]['input']+"\n"+"output:"
else:
if generic_instruction!='': prompt=generic_instruction+"\n"+'input: '+null_word+"\n"+"output:"
else: prompt='input: '+null_word+"\n"+"output:"
if 'Completion' in labels[0]:
prompt = prompt + ' Completion'
promptlist.append(prompt)
answer = data['Instances'][indexlist[i]]['output'][0].strip(".").replace('Completion ', '')
answer = one_token(answer)
answerlist.append(answer)
return promptlist, answerlist, indexlist
def training_encodeinstruction (task, instruction_structure =['Definition','Prompt','Things to Avoid','Emphasis & Caution', 'Negative Examples Full Explanations', 'Positive Examples Full Explanations'], number_of_examples=0, number_of_instances= 100, null_word=None, seed=0, modified={}):
random.seed(0)
with open('data/ExpandedNaturalInstructions/'+task) as json_file:
data = json.load(json_file)
labels = list(set([data["Instances"][i]["output"][0] for i in range(len(data["Instances"])) ]))
labels.sort()
assert len(labels) < 25, "Check {} is a classification task.".format(task)
instances_per_label = number_of_instances // len(labels)
remainder = number_of_instances % len(labels)
instance_pools = {label:{'indices':[]} for label in labels}
for i, inst in enumerate(data["Instances"]):
label = inst['output'][0]
instance_pools[label]['indices'].append(i)
remaining = 0
test_pools = {}
for l, label in enumerate(labels):
if len(instance_pools[label]['indices']) >= 4 + instances_per_label: #see comment in function above
num = instances_per_label
if l < remainder: num += 1
test_pools[label] = random.sample(instance_pools[label]['indices'], num)
instance_pools[label]['indices'] = [i for i in instance_pools[label]['indices'] if i not in test_pools[label]]
else:
num = len(instance_pools[label]['indices']) - 4
remaining += instances_per_label - num
test_pools[label] = random.sample(instance_pools[label]['indices'], num)
instance_pools[label]['indices'] = [i for i in instance_pools[label]['indices'] if i not in test_pools[label]]
all_remaining_indices = []
remaining = number_of_instances - sum([len(t) for t in test_pools.values()])
for label in labels: all_remaining_indices.extend(instance_pools[label]['indices'])
remaining_test = random.sample(all_remaining_indices, remaining)
for t in remaining_test:
label = data['Instances'][t]['output'][0]
test_pools[label].append(t)
instance_pools[label]['indices'].remove(t)
indexlist = []
for label in labels: indexlist.extend(test_pools[label])
assert len(indexlist) == number_of_instances, pdb.set_trace()
random.seed(seed)
if number_of_examples == -1: total_num_examples = 1
else: total_num_examples = number_of_examples * len(labels)
pos_examples = {label:[] for label in labels}
for eg in data["Positive Examples"]:
label = eg['output']
pos_examples[label].append(eg)
for label in labels:
for id in instance_pools[label]['indices']:
inst = data["Instances"][id]
inst['output'] = inst['output'][0]
pos_examples[label].append(inst)
chosen_examples = []
if number_of_examples > 0 :
for label in labels: chosen_examples.extend(random.sample(pos_examples[label], number_of_examples))
elif number_of_examples == -1:
label = random.sample(labels, 1)
chosen_examples.extend(random.sample(pos_examples[label], number_of_examples))
assert len(chosen_examples) == total_num_examples
random.shuffle(chosen_examples)
train_indexlist = list(range(len(data['Instances'])))
train_indexlist = [i for i in train_indexlist if i not in indexlist and data['Instances'][i] not in chosen_examples]
dev_len = round(0.1*len(train_indexlist))
dev_indexlist = random.sample(train_indexlist, dev_len)
train_indexlist = [i for i in train_indexlist if i not in dev_indexlist]
generic_instruction=''
for i in instruction_structure:
if i!='Positive Examples Full Only' and i!='Positive Examples Full Explanations' and i!='Negative Examples Full Explanations':
if data[i]!='-':
if i in modified.keys():
data[i] = modified[i]
data[i] = data[i].replace('\n' + 'Things to avoid: -', '')
data[i] = data[i].replace('\n' + 'Emphasis & Caution: -', '')
# pdb.set_trace()
if generic_instruction=='':
generic_instruction=generic_instruction+i+': '+data[i].strip()
else:
generic_instruction=generic_instruction+"\n"+i+': '+data[i].strip()
elif i=='Positive Examples Full Only' :
for j in range(total_num_examples):
if generic_instruction!='':
generic_instruction=generic_instruction+"\n"+'input: '+chosen_examples[j]['input'] + "\n"+ 'output: '+ one_token(chosen_examples[j]['output'])
else:
generic_instruction=generic_instruction+'input: '+chosen_examples[j]['input'] + "\n"+ 'output: '+one_token(chosen_examples[j]['output'])
elif i=='Positive Examples Full Explanations' : #This mode of Natural Instructions not supported
assert False
elif i=='Negative Examples Full Explanations' : #This mode of Natural Instructions not supported
assert False
promptlist=[]
answerlist=[]
for i in range(number_of_instances):
if null_word is None:
if generic_instruction!= '': prompt=generic_instruction+"\n"+'input: '+data['Instances'][indexlist[i]]['input']+"\n"+"output:"
else: prompt='input: '+data['Instances'][indexlist[i]]['input']+"\n"+"output:"
else:
if generic_instruction!='': prompt=generic_instruction+"\n"+'input: '+null_word+"\n"+"output:"
else: prompt='input: '+null_word+"\n"+"output:"
if 'Completion' in labels[0]:
prompt = prompt + ' Completion'
promptlist.append(prompt)
answer = data['Instances'][indexlist[i]]['output'][0].strip(".").replace('Completion ', '')
answer = one_token(answer)
answerlist.append(answer)
train_promptlist=[]
train_answerlist=[]
for i in range(len(train_indexlist)):
if null_word is None:
if generic_instruction!= '': prompt=generic_instruction+"\n"+'input: '+data['Instances'][train_indexlist[i]]['input']+"\n"+"output:"
else: prompt='input: '+data['Instances'][train_indexlist[i]]['input']+"\n"+"output:"
else:
if generic_instruction!='': prompt=generic_instruction+"\n"+'input: '+null_word+"\n"+"output:"
else: prompt='input: '+null_word+"\n"+"output:"
if 'Completion' in labels[0]:
prompt = prompt + ' Completion'
train_promptlist.append(prompt)
train_answer = data['Instances'][train_indexlist[i]]['output'].strip(".").replace('Completion ', '')
train_answer = one_token(train_answer)
train_answerlist.append(train_answer)
dev_promptlist=[]
dev_answerlist=[]
for i in range(len(dev_indexlist)):
if null_word is None:
if generic_instruction!= '': prompt=generic_instruction+"\n"+'input: '+data['Instances'][dev_indexlist[i]]['input']+"\n"+"output:"
else: prompt='input: '+data['Instances'][dev_indexlist[i]]['input']+"\n"+"output:"
else:
if generic_instruction!='': prompt=generic_instruction+"\n"+'input: '+null_word+"\n"+"output:"
else: prompt='input: '+null_word+"\n"+"output:"
if 'Completion' in labels[0]:
prompt = prompt + ' Completion'
dev_promptlist.append(prompt)
dev_answer = data['Instances'][dev_indexlist[i]]['output'].strip(".").replace('Completion ', '')
dev_answer = one_token(dev_answer)
dev_answerlist.append(dev_answer)
return promptlist, answerlist, indexlist, train_promptlist, train_answerlist, train_indexlist, dev_promptlist, dev_answerlist, dev_indexlist