File size: 9,049 Bytes
302920f |
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 |
# Copyright 2024-present the HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import random
import numpy as np
import torch
from datasets import load_dataset
"""
doc https://huggingface.co/docs/datasets/loading
doc https://huggingface.co/docs/datasets/process
doc https://huggingface.co/blog/llama2#how-to-prompt-llama-2
"""
def set_seed(seed):
np.random.seed(seed)
torch.random.manual_seed(seed)
def sample_train_loaders(name, tokenizer, nsamples=128, seed=0, seqlen=2048):
set_seed(seed)
if "wikitext2" in name:
traindata = load_dataset(
"wikitext",
"wikitext-2-raw-v1",
split="train",
)
traindata = "\n\n".join(traindata["text"])
elif "c4" in name:
traindata = load_dataset(
"allenai/c4",
"allenai--c4",
data_files={"train": "en/c4-train.00000-of-01024.json.gz"},
split="train",
)
traindata = "\n\n".join(traindata["text"])
else:
raise NotImplementedError
trainloader = []
for _ in range(nsamples):
i = random.randint(0, len(traindata) - seqlen * 2 - 1)
j = i + seqlen * 2
# breakpoint()
trainenc = tokenizer(traindata[i:j], return_tensors="pt")
inp = trainenc.input_ids[:, :seqlen]
trainloader.append(inp)
return trainloader
def get_redpajama_train(tokenizer, percent=10, seed=3, batch_size=128, max_length=2048):
def tokenization(example):
return tokenizer(example["text"], truncation=True, max_length=max_length)
if percent != 100:
split = f"train[:{int(850000 * percent / 100)}]"
else:
split = "train"
dataset = load_dataset("togethercomputer/RedPajama-Data-1T-Sample", split=split)
processed_dataset = dataset.map(tokenization, batched=True, batch_size=batch_size, num_proc=os.cpu_count())
return processed_dataset
def get_english_quote(dataset_name, tokenizer):
data = load_dataset(dataset_name)
data = data.map(lambda samples: tokenizer(samples["quote"]), batched=True)
return data["train"]
def get_qat_dataset(name, tokenizer, data_percent):
if name == "red_pajama":
data = get_redpajama_train(tokenizer, data_percent)
elif name == "Abirate/english_quotes":
data = get_english_quote(name, tokenizer)
else:
raise NotImplementedError
data = data.shuffle()
return data
llama_chat_format = """<s>[INST] <<SYS>>
"Below is an instruction that describes a task. Write a response that appropriately completes the request."
<</SYS>>
{instruction} [/INST] {response} </s>
"""
def get_calib_data(name, tokenizer, model_id, nsamples, seqlen=2048, seed=3):
print(f" get_data_from: {name}, nsamples={nsamples}, seqlen={seqlen}, {seed}")
cache_file = f"cache/{name}_{model_id.replace('/', '_')}_{nsamples}_{seqlen}_{seed}.pt"
traindataset = []
if not os.path.exists("cache"):
os.makedirs("cache")
if os.path.exists(cache_file):
print(f"found data file: {cache_file}")
traindataset = torch.load(cache_file)
print("loaded ...")
return traindataset
if name == "c4":
traindata = load_dataset(
"allenai/c4",
"allenai--c4",
data_files={"train": "en/c4-train.00000-of-01024.json.gz"},
split="train",
)
tot_text = "\n\n".join(traindata["text"])
elif name == "wikitext2":
traindata = load_dataset("wikitext", "wikitext-2-raw-v1", split="train")
tot_text = "\n\n".join(traindata["text"])
elif name == "ptb":
traindata = load_dataset(
"ptb_text_only",
"penn_treebank",
split="train",
)
tot_text = "\n\n".join(traindata["sentence"])
elif name == "traivia_qa":
traindata = load_dataset("trivia_qa", "rc", split="train")
tot_text = "\n\n".join(traindata["question"])
elif name == "nqopen":
traindata = load_dataset("nq_open", split="train")
tot_text = "\n\n".join(traindata["question"])
elif name == "alpaca":
selected_data_dict = load_dataset("iboing/alpaca_data", split="train").shuffle(seed=seed).take(nsamples)
for example in selected_data_dict:
if example.get("input", "") == "":
s = llama_chat_format.format(instruction=example["instruction"], response=example["output"])
trainenc = tokenizer(s, return_tensors="pt")
inp = trainenc.input_ids[:, :seqlen]
attention_mask = torch.ones_like(inp)
traindataset.append({"input_ids": inp, "attention_mask": attention_mask})
print("example instruction:", s)
torch.save(traindataset, cache_file)
return traindataset
elif name == "MetaMATH":
selected_data_dict = load_dataset("iboing/MetaMathQA-395K", split="train").shuffle(seed=seed).take(nsamples)
for example in selected_data_dict:
if example.get("input", "") == "":
s = llama_chat_format.format(instruction=example["query"], response=example["response"])
trainenc = tokenizer(s, return_tensors="pt")
inp = trainenc.input_ids[:, :seqlen]
attention_mask = torch.ones_like(inp)
traindataset.append({"input_ids": inp, "attention_mask": attention_mask})
print("example instruction:", s)
torch.save(traindataset, cache_file)
return traindataset
elif name == "codefeedback":
selected_data_dict = (
load_dataset("iboing/CodeFeedback-Filtered-Instruction", split="train").shuffle(seed=seed).take(nsamples)
)
for example in selected_data_dict:
if example.get("input", "") == "":
s = llama_chat_format.format(instruction=example["query"], response=example["answer"])
trainenc = tokenizer(s, return_tensors="pt")
inp = trainenc.input_ids[:, :seqlen]
attention_mask = torch.ones_like(inp)
traindataset.append({"input_ids": inp, "attention_mask": attention_mask})
print("example instruction:", s)
torch.save(traindataset, cache_file)
return traindataset
elif name == "WizLMinstruct":
selected_data_dict = (
load_dataset("iboing/WizardLM_evol_instruct_V2_143k", split="train").shuffle(seed=seed).take(nsamples)
)
for example in selected_data_dict:
if example.get("input", "") == "":
s = llama_chat_format.format(
instruction=example["conversation"][0]["human"], response=example["conversation"][0]["assistant"]
)
trainenc = tokenizer(s, return_tensors="pt")
inp = trainenc.input_ids[:, :seqlen]
attention_mask = torch.ones_like(inp)
traindataset.append({"input_ids": inp, "attention_mask": attention_mask})
print("example instruction:", s)
torch.save(traindataset, cache_file)
return traindataset
else:
raise NotImplementedError
print(f"tot_text={len(tot_text)}")
for _ in range(nsamples):
i = random.randint(0, len(tot_text) - seqlen - 1)
j = i + seqlen * 10
trainenc = tokenizer(tot_text[i:j], return_tensors="pt")
inp = trainenc.input_ids[:, :seqlen]
attention_mask = torch.ones_like(inp)
traindataset.append({"input_ids": inp, "attention_mask": attention_mask})
torch.save(traindataset, cache_file)
return traindataset
def get_eval_loaders(name, tokenizer):
if "wikitext2" in name:
testdata = load_dataset(
"wikitext",
"wikitext-2-raw-v1",
split="test",
)
testenc = tokenizer("\n\n".join(testdata["text"]), return_tensors="pt")
return testenc
if "ptb" in name:
valdata = load_dataset(
"ptb_text_only",
"penn_treebank",
split="validation",
)
testenc = tokenizer("\n\n".join(valdata["sentence"]), return_tensors="pt")
return testenc
if "c4" in name:
testdata = load_dataset(
"allenai/c4",
"allenai--c4",
data_files={"validation": "en/c4-validation.00000-of-00008.json.gz"},
split="validation",
)
testenc = tokenizer("\n\n".join(testdata["text"]), return_tensors="pt")
return testenc
raise NotImplementedError
|