Fine-tuning Engine

1from coren import Coren
2
3client = Coren()
4
5# Создание датасета
6dataset = client.finetuning.datasets.create(
7    name="bank-support-dialogues",
8    description="Диалоги службы поддержки банка",
9    format="jsonl",
10    tags=["support", "banking", "v2"]
11)
12
13# Загрузка данных
14client.finetuning.datasets.upload(
15    dataset_id=dataset.id,
16    file="training_data.jsonl"
17)
18
19# Проверка статистики
20stats = client.finetuning.datasets.get(dataset.id)
21print(f"Примеров: {stats.num_examples}")
22print(f"Токенов: {stats.total_tokens}")

1# Запуск LoRA fine-tuning
2job = client.finetuning.jobs.create(
3    name="bank-support-llm-v2",
4    base_model="llama-3-70b",
5    dataset_id=dataset.id,
6    method="lora",  # "lora", "qlora", "full"
7    hyperparameters={
8        "learning_rate": 2e-5,
9        "epochs": 3,
10        "batch_size": 4,
11        "warmup_steps": 100,
12        "lora_rank": 16,
13        "lora_alpha": 32,
14        "lora_dropout": 0.05,
15        "target_modules": ["q_proj", "v_proj", "k_proj", "o_proj"]
16    },
17    compute={
18        "gpu_type": "A100",
19        "num_gpus": 2
20    },
21    validation_split": 0.1,
22    early_stopping={
23        "metric": "eval_loss",
24        "patience": 3
25    }
26)
27
28print(f"Задача создана: {job.id}")
29print(f"Статус: {job.status}")

1import time
2
3# Мониторинг прогресса
4while job.status in ["pending", "running"]:
5    job = client.finetuning.jobs.get(job.id)
6
7    if job.progress:
8        print(f"Эпоха: {job.progress.current_epoch}/{job.progress.total_epochs}")
9        print(f"Шаг: {job.progress.current_step}/{job.progress.total_steps}")
10
11    if job.metrics:
12        print(f"Train Loss: {job.metrics.train_loss:.4f}")
13        print(f"Eval Loss: {job.metrics.eval_loss:.4f}")
14
15    time.sleep(60)
16
17if job.status == "completed":
18    print(f"Обучение завершено!")
19    print(f"Модель: {job.output_model_id}")
20elif job.status == "failed":
21    print(f"Ошибка: {job.error}")

1# Запуск оценки
2evaluation = client.finetuning.evaluations.create(
3    model_id=job.output_model_id,
4    eval_dataset_id="ds_eval_bank_support",
5    metrics=["perplexity", "accuracy", "bleu"],
6    num_samples=500
7)
8
9# Получение результатов
10results = client.finetuning.evaluations.get(evaluation.id)
11
12print("Результаты оценки:")
13print(f"  Perplexity: {results.metrics.perplexity}")
14print(f"  Accuracy: {results.metrics.accuracy}")
15print(f"  BLEU: {results.metrics.bleu}")
16
17print("\nСравнение с базовой моделью:")
18print(f"  Perplexity: {results.comparison.improvement.perplexity}")
19print(f"  Accuracy: {results.comparison.improvement.accuracy}")