pipka

app.py

@@ -1,30 +1,66 @@
-import gradio as gr
-import torch
-from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig, BitsAndBytesConfig
+import datasets
+import evaluate
+import pandas as pd
+import numpy as np
+from datasets import Dataset
+from sklearn.model_selection import train_test_split
+from transformers import (AutoTokenizer, AutoModelForSequenceClassification,
+                          TrainingArguments, Trainer)
 
-model_name = "ai-sage/GigaChat-20B-A3B-instruct-v1.5-int8"
+model_name = "DeepPavlov/rubert-base-cased"
 
-bnb_config = BitsAndBytesConfig(
-    load_in_4bit=True,
-    bnb_4bit_quant_type="nf4",
-    bnb_4bit_compute_dtype=torch.bfloat16,
-    bnb_4bit_use_double_quant=True,
-    llm_int8_enable_fp32_cpu_offload=True
-)
+# Login using e.g. `huggingface-cli login` to access this dataset
+splits = {'train': 'data/train-00000-of-00001.parquet', 'test': 'data/test-00000-of-00001.parquet'}
+df = pd.read_parquet("hf://datasets/mteb/RuSciBenchOECDClassification/" + splits["train"])
 
-tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
-model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, quantization_config=bnb_config, device_map="auto")
-model.generation_config = GenerationConfig.from_pretrained(model_name)
+# Конвертируем датафрейм в Dataset
+train, test = train_test_split(df, test_size=0.2)
+train = Dataset.from_pandas(train)
+test = Dataset.from_pandas(test)
 
-messages = [
-    {"role": "user", "content": "Докажи теорему о неподвижной точке"}
-]
-input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
-outputs = model.generate(input_tensor.to(model.device))
+# Выполняем предобработку текста
+tokenizer = AutoTokenizer.from_pretrained(model_name)
 
-result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=False)
+def tokenize_function(examples):
+	return tokenizer(examples['text'], padding='max_length', truncation=True)
 
-with gr.Blocks() as demo:
-    gr.Textbox(result, label="Вывод строки", interactive=False)
+tokenized_train = train.map(tokenize_function)
+tokenized_test = test.map(tokenize_function)
 
-demo.launch()
+# Загружаем предобученную модель
+model = AutoModelForSequenceClassification.from_pretrained(
+	model_name,
+	num_labels=28)
+
+# Задаем параметры обучения
+training_args = TrainingArguments(
+	output_dir = 'test_trainer_log',
+	evaluation_strategy = 'epoch',
+	per_device_train_batch_size = 6,
+	per_device_eval_batch_size = 6,
+	num_train_epochs = 5,
+	report_to='none')
+
+# Определяем как считать метрику
+metric = evaluate.load('f1')
+def compute_metrics(eval_pred):
+	logits, labels = eval_pred
+	predictions = np.argmax(logits, axis=-1)
+	return metric.compute(predictions=predictions, references=labels)
+
+# Выполняем обучение
+trainer = Trainer(
+	model = model,
+	args = training_args,
+	train_dataset = tokenized_train,
+	eval_dataset = tokenized_test,
+	compute_metrics = compute_metrics)
+
+trainer.train()
+
+# Сохраняем модель
+save_directory = './pt_save_pretrained'
+#tokenizer.save_pretrained(save_directory)
+model.save_pretrained(save_directory)
+#alternatively save the trainer
+#trainer.save_model('CustomModels/CustomHamSpam')