Upload 2 files
Browse files- convert_into_distilbert_dataset.py +120 -0
- fine-tune-distil-bert.ipynb +199 -0
convert_into_distilbert_dataset.py
ADDED
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# The purpose of this file is to take given texts
|
2 |
+
# Put AI ones into negative and human ones into positive
|
3 |
+
# While making sure to split all the texts into word by word
|
4 |
+
# To ensure searching before the text has finished streaming
|
5 |
+
|
6 |
+
# Example this: "The dog walked over the pavement." will be turned into:
|
7 |
+
# The
|
8 |
+
# The dog
|
9 |
+
# The dog walked
|
10 |
+
# The dog walked over
|
11 |
+
# The dog walked over the
|
12 |
+
# The dog walked over the pavement
|
13 |
+
# The dog walked over the pavement.
|
14 |
+
|
15 |
+
# Example data row:
|
16 |
+
# {"query": "Write a story about dogs", "pos": ["lorem ipsum..."], "neg": ["lorem ipsum..."]}
|
17 |
+
|
18 |
+
import re
|
19 |
+
import ujson as json
|
20 |
+
import random
|
21 |
+
from tqdm import tqdm
|
22 |
+
|
23 |
+
def split_string(text):
|
24 |
+
"""Split a given text by spaces and punctuation"""
|
25 |
+
# Split the text by spaces
|
26 |
+
words = text.split()
|
27 |
+
|
28 |
+
# For now we disabled further splitting because of issues
|
29 |
+
# # Further split each word by punctuation using regex
|
30 |
+
# split_words = []
|
31 |
+
# for word in words:
|
32 |
+
# # Find all substrings that match the pattern: either a word or a punctuation mark
|
33 |
+
# split_words.extend(re.findall(r'\w+|[^\w\s]', word))
|
34 |
+
|
35 |
+
return words
|
36 |
+
|
37 |
+
reddit_vs_synth_writing_prompts = []
|
38 |
+
with open("writing_prompts/reddit_vs_synth_writing_prompts.jsonl", "r") as f:
|
39 |
+
temp = f.read()
|
40 |
+
for line in temp.splitlines():
|
41 |
+
loaded_object = json.loads(line)
|
42 |
+
if not "story_human" in loaded_object: # Remove ones where we don't have human data
|
43 |
+
continue
|
44 |
+
|
45 |
+
reddit_vs_synth_writing_prompts.append(loaded_object)
|
46 |
+
|
47 |
+
dataset_entries = []
|
48 |
+
|
49 |
+
SAVE_FILE_NAME = "bert_reddit_vs_synth_writing_prompts.jsonl"
|
50 |
+
|
51 |
+
def add_streamed_data(data):
|
52 |
+
entries = []
|
53 |
+
data_parts = split_string(data)
|
54 |
+
|
55 |
+
for i in range(len(data_parts)):
|
56 |
+
streamed_so_far = " ".join(data_parts[:i + 1]) # Since python slicing is exclusive toward the end
|
57 |
+
entries.append({"text": streamed_so_far, "label": HUMAN_LABEL})
|
58 |
+
|
59 |
+
return entries
|
60 |
+
|
61 |
+
with open(SAVE_FILE_NAME, "w") as f:
|
62 |
+
f.write("")
|
63 |
+
|
64 |
+
NUM_OF_TURNS_TO_DUMP = 200
|
65 |
+
i = 0
|
66 |
+
for data in tqdm(reddit_vs_synth_writing_prompts):
|
67 |
+
# {"text": "AI-generated text example 1", "label": 1},
|
68 |
+
# Assuming 1 means AI generated, 0 means human
|
69 |
+
HUMAN_LABEL = 0
|
70 |
+
AI_LABEL = 1
|
71 |
+
i += 1
|
72 |
+
|
73 |
+
# Below is to enable writing dataset part by part
|
74 |
+
if i == NUM_OF_TURNS_TO_DUMP:
|
75 |
+
i = 0
|
76 |
+
dumped_string = ""
|
77 |
+
dumped_entries = []
|
78 |
+
for entry in dataset_entries:
|
79 |
+
dumped_entries.append(json.dumps(entry))
|
80 |
+
|
81 |
+
dumped_string = "\n".join(dumped_entries) + "\n"
|
82 |
+
|
83 |
+
with open(SAVE_FILE_NAME, "a") as f:
|
84 |
+
f.write(dumped_string)
|
85 |
+
|
86 |
+
dataset_entries = []
|
87 |
+
|
88 |
+
if False: # Disable Streaming
|
89 |
+
# Add streamed data
|
90 |
+
human_entries = add_streamed_data(data["story_human"])
|
91 |
+
dataset_entries.extend(human_entries)
|
92 |
+
|
93 |
+
ai_data = []
|
94 |
+
if data.get("story_opus"):
|
95 |
+
ai_data.extend(add_streamed_data(data["story_opus"]))
|
96 |
+
if data.get("story_gpt_3_5"):
|
97 |
+
ai_data.extend(add_streamed_data(data["story_gpt_3_5"]))
|
98 |
+
|
99 |
+
dataset_entries.extend(ai_data)
|
100 |
+
|
101 |
+
else:
|
102 |
+
# Add without streaming
|
103 |
+
dataset_entries.append({"text": data["story_human"], "label": HUMAN_LABEL})
|
104 |
+
|
105 |
+
ai_data = []
|
106 |
+
if data.get("story_opus"):
|
107 |
+
dataset_entries.append({"text": data["story_opus"], "label": AI_LABEL})
|
108 |
+
if data.get("story_gpt_3_5"):
|
109 |
+
dataset_entries.append({"text": data["story_gpt_3_5"], "label": AI_LABEL})
|
110 |
+
|
111 |
+
# Dump as JSONL
|
112 |
+
dumped_string = ""
|
113 |
+
dumped_entries = []
|
114 |
+
for entry in dataset_entries:
|
115 |
+
dumped_entries.append(json.dumps(entry))
|
116 |
+
|
117 |
+
dumped_string = "\n".join(dumped_entries) + "\n"
|
118 |
+
|
119 |
+
with open(SAVE_FILE_NAME, "a") as f:
|
120 |
+
f.write(dumped_string)
|
fine-tune-distil-bert.ipynb
ADDED
@@ -0,0 +1,199 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {},
|
7 |
+
"outputs": [],
|
8 |
+
"source": [
|
9 |
+
"!pip install torch transformers scikit-learn wandb accelerate tqdm\n",
|
10 |
+
"from IPython.display import clear_output\n",
|
11 |
+
"clear_output(wait=True)\n",
|
12 |
+
"print(\".\")"
|
13 |
+
]
|
14 |
+
},
|
15 |
+
{
|
16 |
+
"cell_type": "code",
|
17 |
+
"execution_count": null,
|
18 |
+
"metadata": {},
|
19 |
+
"outputs": [],
|
20 |
+
"source": [
|
21 |
+
"!apt-get update\n",
|
22 |
+
"!apt-get install zstd\n",
|
23 |
+
"!tar --use-compress-program=unzstd -xvf bert_streamed_dataset.tar.zst\n",
|
24 |
+
"clear_output(wait=True)\n",
|
25 |
+
"print(\".\")"
|
26 |
+
]
|
27 |
+
},
|
28 |
+
{
|
29 |
+
"cell_type": "code",
|
30 |
+
"execution_count": null,
|
31 |
+
"metadata": {},
|
32 |
+
"outputs": [],
|
33 |
+
"source": [
|
34 |
+
"import torch\n",
|
35 |
+
"from transformers import DistilBertTokenizer, DistilBertForSequenceClassification, Trainer, TrainingArguments\n",
|
36 |
+
"from sklearn.model_selection import train_test_split\n",
|
37 |
+
"from tqdm import tqdm\n",
|
38 |
+
"import wandb\n",
|
39 |
+
"import json\n",
|
40 |
+
"\n",
|
41 |
+
"# Initialize W&B\n",
|
42 |
+
"wandb.init(project=\"distilbert-ai-text-classification\")\n",
|
43 |
+
"\n",
|
44 |
+
"# Check if MPS is available and set the device\n",
|
45 |
+
"device = torch.device(\"cuda\") if torch.cuda.is_available() else torch.device(\"cpu\")\n",
|
46 |
+
"print(device)\n",
|
47 |
+
"\n",
|
48 |
+
"# Load pre-trained DistilBERT tokenizer and model\n",
|
49 |
+
"tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')\n",
|
50 |
+
"model = DistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased', num_labels=2)\n",
|
51 |
+
"model.to(device)"
|
52 |
+
]
|
53 |
+
},
|
54 |
+
{
|
55 |
+
"cell_type": "code",
|
56 |
+
"execution_count": null,
|
57 |
+
"metadata": {},
|
58 |
+
"outputs": [],
|
59 |
+
"source": [
|
60 |
+
"# Load the JSONL dataset\n",
|
61 |
+
"data = []\n",
|
62 |
+
"total_num_of_lines = 0\n",
|
63 |
+
"with open('bert_reddit_vs_synth_writing_prompts.jsonl', 'r') as infile:\n",
|
64 |
+
" for line in tqdm(infile, desc=\"Checking dataset size\"):\n",
|
65 |
+
" total_num_of_lines += 1\n",
|
66 |
+
"\n",
|
67 |
+
"with open('bert_reddit_vs_synth_writing_prompts.jsonl', 'r') as infile:\n",
|
68 |
+
" for line in tqdm(infile, desc=\"Loading dataset\", total=total_num_of_lines):\n",
|
69 |
+
" data.append(json.loads(line))\n",
|
70 |
+
"\n",
|
71 |
+
"# Extract texts and labels\n",
|
72 |
+
"print(\"Extracting texts and labels\")\n",
|
73 |
+
"texts = [entry['text'] for entry in data]\n",
|
74 |
+
"labels = [entry['label'] for entry in data]\n",
|
75 |
+
"\n",
|
76 |
+
"# Tokenize the text\n",
|
77 |
+
"print(\"Tokenizing text\")\n",
|
78 |
+
"inputs = tokenizer(texts, padding=True, truncation=True, return_tensors=\"pt\")\n",
|
79 |
+
"\n",
|
80 |
+
"# Move input tensors to the device\n",
|
81 |
+
"print(\"Moving input tensors\")\n",
|
82 |
+
"inputs = {key: val for key, val in inputs.items()}\n",
|
83 |
+
"\n",
|
84 |
+
"# Split the data into training and validation sets\n",
|
85 |
+
"print(\"Splitting data into train and validation\")\n",
|
86 |
+
"train_inputs, val_inputs, train_labels, val_labels = train_test_split(\n",
|
87 |
+
" inputs['input_ids'], labels, test_size=0.2, random_state=42)\n",
|
88 |
+
"\n",
|
89 |
+
"train_attention_masks, val_attention_masks, _, _ = train_test_split(\n",
|
90 |
+
" inputs['attention_mask'], labels, test_size=0.2, random_state=42)\n",
|
91 |
+
"\n",
|
92 |
+
"# Create a PyTorch dataset\n",
|
93 |
+
"class TextDataset(torch.utils.data.Dataset):\n",
|
94 |
+
" def __init__(self, input_ids, attention_masks, labels):\n",
|
95 |
+
" self.input_ids = input_ids\n",
|
96 |
+
" self.attention_masks = attention_masks\n",
|
97 |
+
" self.labels = labels\n",
|
98 |
+
"\n",
|
99 |
+
" def __len__(self):\n",
|
100 |
+
" return len(self.labels)\n",
|
101 |
+
"\n",
|
102 |
+
" def __getitem__(self, idx):\n",
|
103 |
+
" return {\n",
|
104 |
+
" 'input_ids': self.input_ids[idx],\n",
|
105 |
+
" 'attention_mask': self.attention_masks[idx],\n",
|
106 |
+
" 'labels': torch.tensor(self.labels[idx])\n",
|
107 |
+
" }\n",
|
108 |
+
"\n",
|
109 |
+
"print(\"Creating pytorch datasets\")\n",
|
110 |
+
"train_dataset = TextDataset(train_inputs, train_attention_masks, train_labels)\n",
|
111 |
+
"val_dataset = TextDataset(val_inputs, val_attention_masks, val_labels)"
|
112 |
+
]
|
113 |
+
},
|
114 |
+
{
|
115 |
+
"cell_type": "code",
|
116 |
+
"execution_count": null,
|
117 |
+
"metadata": {},
|
118 |
+
"outputs": [],
|
119 |
+
"source": [
|
120 |
+
"# Reduce eval set to X examples to speed up training\n",
|
121 |
+
"NUM_OF_EVAL_EXAMPLES = 1000\n",
|
122 |
+
"val_inputs_subset = val_inputs[:NUM_OF_EVAL_EXAMPLES]\n",
|
123 |
+
"val_attention_masks_subset = val_attention_masks[:NUM_OF_EVAL_EXAMPLES]\n",
|
124 |
+
"val_labels_subset = val_labels[:NUM_OF_EVAL_EXAMPLES]\n",
|
125 |
+
"\n",
|
126 |
+
"# Create a TextDataset with only X examples\n",
|
127 |
+
"val_dataset = Textdataset(val_inputs_subset, val_attention_masks_subset, val_labels_subset)"
|
128 |
+
]
|
129 |
+
},
|
130 |
+
{
|
131 |
+
"cell_type": "code",
|
132 |
+
"execution_count": null,
|
133 |
+
"metadata": {},
|
134 |
+
"outputs": [],
|
135 |
+
"source": [
|
136 |
+
"# Define the training arguments\n",
|
137 |
+
"training_args = TrainingArguments(\n",
|
138 |
+
" output_dir='./distil-bert-train-results', \n",
|
139 |
+
" num_train_epochs=3, \n",
|
140 |
+
" per_device_train_batch_size=16, \n",
|
141 |
+
" per_device_eval_batch_size=16, \n",
|
142 |
+
" warmup_steps=500, # number of warmup steps for learning rate scheduler\n",
|
143 |
+
" weight_decay=0.01, \n",
|
144 |
+
" logging_dir='./logs', \n",
|
145 |
+
" logging_steps=10, \n",
|
146 |
+
" report_to=\"wandb\", \n",
|
147 |
+
" evaluation_strategy=\"steps\", # Evaluate every logging step\n",
|
148 |
+
" eval_steps=100, # Evaluate every 10 steps\n",
|
149 |
+
" fp16=True,\n",
|
150 |
+
")\n",
|
151 |
+
"\n",
|
152 |
+
"# Create the Trainer\n",
|
153 |
+
"trainer = Trainer(\n",
|
154 |
+
" model=model, # the instantiated 🤗 Transformers model to be trained\n",
|
155 |
+
" args=training_args, # training arguments, defined above\n",
|
156 |
+
" train_dataset=train_dataset, # training dataset\n",
|
157 |
+
" eval_dataset=val_dataset # evaluation dataset\n",
|
158 |
+
")\n",
|
159 |
+
"\n",
|
160 |
+
"# Train the model\n",
|
161 |
+
"trainer.train()\n",
|
162 |
+
"\n",
|
163 |
+
"# Save the model\n",
|
164 |
+
"model.save_pretrained('./distil-bert-train-final-result')\n",
|
165 |
+
"\n",
|
166 |
+
"# Finish the W&B run\n",
|
167 |
+
"wandb.finish()"
|
168 |
+
]
|
169 |
+
},
|
170 |
+
{
|
171 |
+
"cell_type": "code",
|
172 |
+
"execution_count": null,
|
173 |
+
"metadata": {},
|
174 |
+
"outputs": [],
|
175 |
+
"source": []
|
176 |
+
}
|
177 |
+
],
|
178 |
+
"metadata": {
|
179 |
+
"kernelspec": {
|
180 |
+
"display_name": "Python 3 (ipykernel)",
|
181 |
+
"language": "python",
|
182 |
+
"name": "python3"
|
183 |
+
},
|
184 |
+
"language_info": {
|
185 |
+
"codemirror_mode": {
|
186 |
+
"name": "ipython",
|
187 |
+
"version": 3
|
188 |
+
},
|
189 |
+
"file_extension": ".py",
|
190 |
+
"mimetype": "text/x-python",
|
191 |
+
"name": "python",
|
192 |
+
"nbconvert_exporter": "python",
|
193 |
+
"pygments_lexer": "ipython3",
|
194 |
+
"version": "3.10.12"
|
195 |
+
}
|
196 |
+
},
|
197 |
+
"nbformat": 4,
|
198 |
+
"nbformat_minor": 4
|
199 |
+
}
|