File size: 5,498 Bytes
992b12b 11c619a 992b12b 11c619a c9029e6 11c619a a5d4a06 c9029e6 11c619a 992b12b 11c619a 992b12b 11c619a 992b12b |
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 |
import gc
from datasets import load_dataset
from litdata import optimize, TokensLoader
from litgpt.tokenizer import Tokenizer
from functools import partial
def batch_iterator(name=None):
if name in (None, 'Replete-AI/Everything_Instruct_Multilingual'):
dataset = load_dataset('Replete-AI/Everything_Instruct_Multilingual', split='train')
for row in dataset:
text = []
if row['instruction']:
text.append(
'<|im_start|>system\n'
f"{row['instruction']}<|im_end|>"
)
if row['input']:
text.append(
'<|im_start|>user\n'
f"{row['input']}<|im_end|>"
)
if row['output']:
text.append(
'<|im_start|>assistant\n'
f"{row['output']}<|im_end|>"
)
text = '\n'.join(text) + '\n'
yield text
del dataset
gc.collect()
if name in (None, 'HuggingFaceH4/ultrachat_200k'):
dataset = load_dataset('HuggingFaceH4/ultrachat_200k', split='train_sft')
for row in dataset:
text = [
f"<|im_start|>{n['role']}\n{n['content']}<|im_end|>"
for n in row['messages']
]
text = '\n'.join(text) + '\n'
yield text
del dataset
gc.collect()
if name in (None, 'HuggingFaceH4/no_robots'):
dataset = load_dataset('HuggingFaceH4/no_robots', split='train')
for row in dataset:
text = [
f"<|im_start|>{n['role']}\n{n['content']}<|im_end|>"
for n in row['messages']
]
text = '\n'.join(text) + '\n'
yield text
del dataset
gc.collect()
if name in (None, 'datatab/ultrachat_200k_serbian'):
dataset = load_dataset('datatab/ultrachat_200k_serbian', split='train')
for row in dataset:
text = [
f"<|im_start|>{n['role']}\n{n['content']}<|im_end|>"
for n in row['messages_srb']
]
text = '\n'.join(text) + '\n'
yield text
del dataset
gc.collect()
if name in (None, 'datatab/ultrafeedback_binarized_serbian'):
dataset = load_dataset('datatab/ultrafeedback_binarized_serbian', split='train_sft')
for row in dataset:
text = [
f"<|im_start|>{n['role']}\n{n['content']}<|im_end|>"
for n in row['chosen']
]
text = '\n'.join(text) + '\n'
yield text
del dataset
gc.collect()
if name in (None, 'datatab/alpaca-cleaned-serbian-full'):
dataset = load_dataset('datatab/alpaca-cleaned-serbian-full', split='train')
for row in dataset:
text = []
if row['instruction']:
text.append(
'<|im_start|>system\n'
f"{row['instruction']}<|im_end|>"
)
if row['input']:
text.append(
'<|im_start|>user\n'
f"{row['input']}<|im_end|>"
)
if row['output']:
text.append(
'<|im_start|>assistant\n'
f"{row['output']}<|im_end|>"
)
text = '\n'.join(text) + '\n'
yield text
del dataset
gc.collect()
if name in (None, 'datatab/orca_math_world_problem_200k_serbian'):
dataset = load_dataset('datatab/orca_math_world_problem_200k_serbian', split='train')
for row in dataset:
text = []
text.append(
'<|im_start|>user\n'
f"{row['question_translated_srb']}<|im_end|>"
)
text.append(
'<|im_start|>assistant\n'
f"{row['answer_translated_srb']}<|im_end|>"
)
text = '\n'.join(text) + '\n'
yield text
del dataset
gc.collect()
if name in (None, 'datatab/open-orca-slim-serbian'):
dataset = load_dataset('datatab/open-orca-slim-serbian', split='train')
role_map = {'system': 'system', 'human': 'user', 'gpt': 'assistant'}
for row in dataset['conversations']:
text = [
f"<|im_start|>{role_map[n['from']]}\n{n['value']}<|im_end|>"
for n in row
if n
]
text = '\n'.join(text) + '\n'
yield text
del dataset
gc.collect()
def tokenize_fn(dataset_name, tokenizer=None):
for text in batch_iterator(dataset_name):
text_ids = tokenizer.encode(text, bos=False, eos=True)
yield text_ids
datasets_names = [
'Replete-AI/Everything_Instruct_Multilingual',
'HuggingFaceH4/ultrachat_200k',
'HuggingFaceH4/no_robots',
'datatab/ultrachat_200k_serbian',
'datatab/ultrafeedback_binarized_serbian',
'datatab/alpaca-cleaned-serbian-full',
'datatab/orca_math_world_problem_200k_serbian',
'datatab/open-orca-slim-serbian',
]
outputs = optimize(
fn=partial(tokenize_fn, tokenizer=Tokenizer('..')),
inputs=datasets_names,
output_dir='../data/',
# Number of tokens to store by chunks. This is roughly 64MB of tokens per chunk.
chunk_size=((32768 + 1) * 500),
num_workers=16,
)
|