Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,329 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer, TrainerCallback, DataCollatorWithPadding, DefaultDataCollator
|
3 |
+
from openai import OpenAI
|
4 |
+
from huggingface_hub import login
|
5 |
+
import datasets
|
6 |
+
from datasets import Dataset
|
7 |
+
import json
|
8 |
+
import pandas as pd
|
9 |
+
import torch
|
10 |
+
import wandb
|
11 |
+
import os
|
12 |
+
import sys
|
13 |
+
from peft import LoraConfig, TaskType, get_peft_model, AutoPeftModelForCausalLM
|
14 |
+
from sklearn.model_selection import train_test_split
|
15 |
+
|
16 |
+
IS_COLAB = False
|
17 |
+
if "google.colab" in sys.modules or "google.colab" in os.environ:
|
18 |
+
IS_COLAB = True
|
19 |
+
|
20 |
+
# Load env secrets
|
21 |
+
if IS_COLAB:
|
22 |
+
from google.colab import userdata
|
23 |
+
OPENAI_API_KEY=userdata.get('OPENAI_API_KEY')
|
24 |
+
WANDB_API_KEY=userdata.get('WANDB_API_KEY')
|
25 |
+
else:
|
26 |
+
OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY")
|
27 |
+
WANDB_API_KEY = os.environ.get("WANDB_API_KEY")
|
28 |
+
|
29 |
+
# Authenticate Weights and Biases
|
30 |
+
wandb.login(key=WANDB_API_KEY)
|
31 |
+
|
32 |
+
# Custom callback to capture logs
|
33 |
+
class LoggingCallback(TrainerCallback):
|
34 |
+
def __init__(self):
|
35 |
+
self.logs = [] # Store logs
|
36 |
+
|
37 |
+
def on_log(self, args, state, control, logs=None, **kwargs):
|
38 |
+
if logs:
|
39 |
+
self.logs.append(logs) # Append logs to list
|
40 |
+
|
41 |
+
|
42 |
+
class LLMTrainingApp:
|
43 |
+
def __init__(self):
|
44 |
+
# self.metric = datasets.load_metric('sacrebleu')
|
45 |
+
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
46 |
+
self.finetuning_dataset = []
|
47 |
+
self.prompt_template = """### Question: {question} ### Answer: """
|
48 |
+
self.training_output = "/content/peft-model" if IS_COLAB else "./peft-model"
|
49 |
+
self.localpath = "/content/finetuned-model" if IS_COLAB else "./finetuned-model"
|
50 |
+
self.tokenizer = None
|
51 |
+
self.model = None
|
52 |
+
self.model_name = None
|
53 |
+
self.fine_tuned_model = None
|
54 |
+
self.teacher_model = OpenAI(api_key=OPENAI_API_KEY)
|
55 |
+
self.base_models = {
|
56 |
+
"SmolLM": {"hf_name":"HuggingFaceTB/SmolLM2-135M",
|
57 |
+
"model_size": "135M",
|
58 |
+
"training_size": "2T",
|
59 |
+
"context_window": "8192"},
|
60 |
+
"GPT2": {"hf_name":"openai-community/gpt2",
|
61 |
+
"model_size": "137M",
|
62 |
+
"training_size": "2T",
|
63 |
+
"context_window": "1024"}
|
64 |
+
}
|
65 |
+
self.peft_config = LoraConfig(task_type=TaskType.CAUSAL_LM, inference_mode=False, r=8, lora_alpha=32, lora_dropout=0.1)
|
66 |
+
self.logging_callback = LoggingCallback()
|
67 |
+
|
68 |
+
def login_into_hf(self, token):
|
69 |
+
if not token:
|
70 |
+
return "β Please enter a valid token."
|
71 |
+
try:
|
72 |
+
login(token)
|
73 |
+
return f"β
Logged in successfully!"
|
74 |
+
except Exception as e:
|
75 |
+
return f"β Login failed: {str(e)}"
|
76 |
+
|
77 |
+
def select_model(self, model_name):
|
78 |
+
self.model_name = model_name
|
79 |
+
model_hf_name = self.base_models[model_name]["hf_name"]
|
80 |
+
try:
|
81 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_hf_name)
|
82 |
+
self.tokenizer.pad_token = self.tokenizer.eos_token
|
83 |
+
base_model = AutoModelForCausalLM.from_pretrained(
|
84 |
+
model_hf_name,
|
85 |
+
torch_dtype="auto",
|
86 |
+
device_map="auto"
|
87 |
+
)
|
88 |
+
self.model = get_peft_model(base_model, self.peft_config)
|
89 |
+
params = self.model.get_nb_trainable_parameters()
|
90 |
+
percent_trainable = round(100 * (params[0] / params[1]), 2)
|
91 |
+
return f"β
Loaded model into memory! Base Model card: {json.dumps(self.base_models[model_name])} - % of trainable parameters for PEFT model: {percent_trainable}"
|
92 |
+
except Exception as e:
|
93 |
+
return f"β Failed to load model and/or tokenizer: {str(e)}"
|
94 |
+
|
95 |
+
def create_golden_dataset(self, dataset):
|
96 |
+
try:
|
97 |
+
dataset = pd.DataFrame(dataset)
|
98 |
+
for i, row in dataset.iterrows():
|
99 |
+
self.finetuning_dataset.append({"question": self.prompt_template.format(question=row["Question"]), "answer": row["Answer"]})
|
100 |
+
return "β
Golden dataset created!"
|
101 |
+
except Exception as e:
|
102 |
+
return f"β Failed to create dataset: {str(e)}"
|
103 |
+
|
104 |
+
def extend_dataset(self):
|
105 |
+
try:
|
106 |
+
completion = self.teacher_model.chat.completions.create(
|
107 |
+
model="gpt-4o",
|
108 |
+
messages=[
|
109 |
+
{
|
110 |
+
"role": "user",
|
111 |
+
"content": """Given the following question-answer pairs, generate 10 similar pairs in the following json format below. Do not respond with anything other than the json.
|
112 |
+
```json
|
113 |
+
[
|
114 |
+
{
|
115 |
+
"question": "question 1",
|
116 |
+
"answer": "answer 1"
|
117 |
+
},
|
118 |
+
{
|
119 |
+
"question": "question 2",
|
120 |
+
"answer": "answer 2"
|
121 |
+
}
|
122 |
+
]
|
123 |
+
"""
|
124 |
+
}
|
125 |
+
]
|
126 |
+
)
|
127 |
+
response = completion.choices[0].message.content
|
128 |
+
print(f"raw response: {response}")
|
129 |
+
clean_response = response.replace("```json", "").replace("```", "").strip()
|
130 |
+
print(f"clean response: {clean_response}")
|
131 |
+
new_data = json.loads(clean_response)
|
132 |
+
for i, row in enumerate(new_data):
|
133 |
+
self.finetuning_dataset.append({"question": self.prompt_template.format(question=row["question"]), "answer": row["answer"]})
|
134 |
+
# create df to display
|
135 |
+
df = pd.DataFrame(new_data)
|
136 |
+
return "β
Synthetic dataset generated!", df
|
137 |
+
except Exception as e:
|
138 |
+
return f"β Failed to generate synthetic dataset: {str(e)}", pd.DataFrame()
|
139 |
+
|
140 |
+
def tokenize_function(self, examples):
|
141 |
+
try:
|
142 |
+
# Tokenize the question and answer as input and target (labels) for causal LM
|
143 |
+
encoding = self.tokenizer(examples['question'], examples['answer'], padding=True)
|
144 |
+
# Set the labels as the input_ids
|
145 |
+
encoding['labels'] = encoding['input_ids'].copy()
|
146 |
+
return encoding
|
147 |
+
except Exception as e:
|
148 |
+
return f"β Failed to tokenize input: {str(e)}"
|
149 |
+
|
150 |
+
|
151 |
+
def prepare_data_for_training(self):
|
152 |
+
try:
|
153 |
+
dataset = Dataset.from_dict({
|
154 |
+
"question": [entry["question"] for entry in self.finetuning_dataset],
|
155 |
+
"answer": [entry["answer"] for entry in self.finetuning_dataset],
|
156 |
+
})
|
157 |
+
dataset = dataset.map(self.tokenize_function, batched=True)
|
158 |
+
train_dataset, test_dataset = dataset.train_test_split(test_size=0.2).values()
|
159 |
+
return {"train": train_dataset, "test": test_dataset}
|
160 |
+
except Exception as e:
|
161 |
+
return f"β Failed to prepare data for training: {str(e)}"
|
162 |
+
|
163 |
+
|
164 |
+
def compute_bleu(self, eval_pred):
|
165 |
+
predictions, labels = eval_pred
|
166 |
+
# # Flatten predictions and labels if they are in nested lists
|
167 |
+
# predictions = predictions.flatten()
|
168 |
+
# labels = labels.flatten()
|
169 |
+
# # Ensure that predictions and labels are integers
|
170 |
+
# predictions = predictions.astype(int) # Convert to integer
|
171 |
+
# labels = labels.astype(int) # Convert to integer
|
172 |
+
# # Decode the predicted tokens
|
173 |
+
# decoded_preds = self.tokenizer.decode(predictions, skip_special_tokens=True)
|
174 |
+
# decoded_labels = self.tokenizer.decode(labels, skip_special_tokens=True)
|
175 |
+
# result = self.metric.compute(predictions=[decoded_preds], references=[[decoded_labels]])
|
176 |
+
result = {"bleu": 1}
|
177 |
+
return result
|
178 |
+
|
179 |
+
def log_generator(self):
|
180 |
+
""" Continuously send logs to frontend during training """
|
181 |
+
for log in self.logging_callback.logs:
|
182 |
+
yield str(log)
|
183 |
+
|
184 |
+
def train_model(self):
|
185 |
+
try:
|
186 |
+
tokenized_datasets = self.prepare_data_for_training()
|
187 |
+
|
188 |
+
# Create training arguments
|
189 |
+
training_args = TrainingArguments(
|
190 |
+
output_dir=self.training_output,
|
191 |
+
learning_rate=1e-3,
|
192 |
+
per_device_train_batch_size=32,
|
193 |
+
per_device_eval_batch_size=32,
|
194 |
+
num_train_epochs=2,
|
195 |
+
weight_decay=0.01,
|
196 |
+
eval_strategy="epoch",
|
197 |
+
save_strategy="epoch",
|
198 |
+
load_best_model_at_end=True,
|
199 |
+
)
|
200 |
+
|
201 |
+
# Create trainer & attach logging callback
|
202 |
+
trainer = Trainer(
|
203 |
+
model=self.model,
|
204 |
+
args=training_args,
|
205 |
+
train_dataset=tokenized_datasets["train"],
|
206 |
+
eval_dataset=tokenized_datasets["test"],
|
207 |
+
tokenizer=self.tokenizer,
|
208 |
+
data_collator=DefaultDataCollator(),
|
209 |
+
compute_metrics=self.compute_bleu,
|
210 |
+
callbacks=[self.logging_callback],
|
211 |
+
)
|
212 |
+
|
213 |
+
# Start training and yield logs in real-time
|
214 |
+
trainer.train()
|
215 |
+
|
216 |
+
# for log in logging_callback.logs:
|
217 |
+
# yield str(log)
|
218 |
+
|
219 |
+
# Save trained model to HF
|
220 |
+
self.model.save_pretrained(self.localpath) # save to local
|
221 |
+
self.model.push_to_hub(f"{self.model_name}-lora")
|
222 |
+
|
223 |
+
return "β
Training complete!"
|
224 |
+
except Exception as e:
|
225 |
+
return f"β Training failed: {str(e)}"
|
226 |
+
|
227 |
+
def run_inference(self, prompt):
|
228 |
+
try:
|
229 |
+
# Load fine-tuned memory into memory and set mode to eval
|
230 |
+
self.fine_tuned_model = AutoPeftModelForCausalLM.from_pretrained(self.localpath)
|
231 |
+
self.fine_tuned_model = self.fine_tuned_model.to(self.device)
|
232 |
+
self.fine_tuned_model.eval()
|
233 |
+
|
234 |
+
# Tokenize input with padding and attention mask
|
235 |
+
inputs = self.tokenizer(prompt, return_tensors="pt", padding=True).to(self.device)
|
236 |
+
|
237 |
+
# Generate response
|
238 |
+
output = self.fine_tuned_model.generate(
|
239 |
+
**inputs,
|
240 |
+
max_length=50, # Limit response length
|
241 |
+
num_return_sequences=1, # Single response
|
242 |
+
temperature=0.7, # Sampling randomness
|
243 |
+
top_p=0.9 # Nucleus sampling
|
244 |
+
)
|
245 |
+
|
246 |
+
response = self.tokenizer.batch_decode(output.detach().cpu().numpy(), skip_special_tokens=True)[0]
|
247 |
+
return response
|
248 |
+
except Exception as e:
|
249 |
+
return f"β Inference failed: {str(e)}"
|
250 |
+
|
251 |
+
def build_ui(self):
|
252 |
+
with gr.Blocks() as demo:
|
253 |
+
gr.Markdown("# LLM Fine-tuning")
|
254 |
+
|
255 |
+
# Model Selection
|
256 |
+
with gr.Group():
|
257 |
+
gr.Markdown("### 1. Login into Hugging Face")
|
258 |
+
with gr.Column():
|
259 |
+
token = gr.Textbox(label="Enter Hugging Face Access Token (w/ write permissions)", type="password")
|
260 |
+
inference_btn = gr.Button("Login", variant="primary")
|
261 |
+
status = gr.Textbox(label="Status")
|
262 |
+
inference_btn.click(self.login_into_hf, inputs=token, outputs=status)
|
263 |
+
|
264 |
+
# Model Selection
|
265 |
+
with gr.Group():
|
266 |
+
gr.Markdown("### 2. Select Model")
|
267 |
+
with gr.Column():
|
268 |
+
model_dropdown = gr.Dropdown([key for key in self.base_models.keys()], label="Small Models")
|
269 |
+
select_model_btn = gr.Button("Select", variant="primary")
|
270 |
+
selected_model_text = gr.Textbox(label="Model Status")
|
271 |
+
select_model_btn.click(self.select_model, inputs=model_dropdown, outputs=[selected_model_text])
|
272 |
+
|
273 |
+
# Create Golden Dataset
|
274 |
+
with gr.Group():
|
275 |
+
gr.Markdown("### 3. Create Golden Dataset")
|
276 |
+
with gr.Column():
|
277 |
+
dataset_table = gr.Dataframe(
|
278 |
+
headers=["Question", "Answer"],
|
279 |
+
value=[["", ""] for _ in range(3)],
|
280 |
+
label="Golden Dataset"
|
281 |
+
)
|
282 |
+
create_data_btn = gr.Button("Create Dataset", variant="primary")
|
283 |
+
dataset_status = gr.Textbox(label="Dataset Status")
|
284 |
+
create_data_btn.click(self.create_golden_dataset, inputs=dataset_table, outputs=[dataset_status])
|
285 |
+
|
286 |
+
# Generate Full Dataset
|
287 |
+
with gr.Group():
|
288 |
+
gr.Markdown("### 4. Extend Dataset with Synthetic Data")
|
289 |
+
with gr.Column():
|
290 |
+
dataset_table = gr.Dataframe(
|
291 |
+
headers=["Question", "Answer"],
|
292 |
+
label="Golden + Synthetic Dataset"
|
293 |
+
)
|
294 |
+
generate_status = gr.Textbox(label="Dataset Generation Status")
|
295 |
+
generate_data_btn = gr.Button("Generate Dataset", variant="primary")
|
296 |
+
generate_data_btn.click(self.extend_dataset, outputs=[generate_status, dataset_table])
|
297 |
+
|
298 |
+
# Train Model & Visualize Loss
|
299 |
+
with gr.Group():
|
300 |
+
gr.Markdown("### 5. Start Logging")
|
301 |
+
with gr.Column():
|
302 |
+
train_status = gr.Textbox(label="Training Status", lines=10)
|
303 |
+
train_btn = gr.Button("Train", variant="primary")
|
304 |
+
train_btn.click(self.log_generator, outputs=[train_status])
|
305 |
+
|
306 |
+
# Train Model & Visualize Loss
|
307 |
+
with gr.Group():
|
308 |
+
gr.Markdown("### 6. Train Model")
|
309 |
+
with gr.Column():
|
310 |
+
train_status = gr.Textbox(label="Training Status")
|
311 |
+
train_btn = gr.Button("Train", variant="primary")
|
312 |
+
train_btn.click(self.train_model, outputs=[train_status])
|
313 |
+
|
314 |
+
# Run Inference
|
315 |
+
with gr.Group():
|
316 |
+
gr.Markdown("### 7. Run Inference")
|
317 |
+
with gr.Column():
|
318 |
+
user_prompt = gr.Textbox(label="Enter Prompt")
|
319 |
+
inference_btn = gr.Button("Run Inference", variant="primary")
|
320 |
+
inference_output = gr.Textbox(label="Inference Output")
|
321 |
+
inference_btn.click(self.run_inference, inputs=user_prompt, outputs=inference_output)
|
322 |
+
|
323 |
+
return demo
|
324 |
+
|
325 |
+
# Create an instance of the app
|
326 |
+
app = LLMTrainingApp()
|
327 |
+
|
328 |
+
# Launch the Gradio app using the class method
|
329 |
+
app.build_ui().launch()
|