from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training from datasets import load_dataset from trl import SFTConfig, SFTTrainer import torch # 1. Configurações MODEL_NAME = "Qwen/Qwen2.5-Coder-7B-Instruct" # 2. Carregar Tokenizer e Modelo (Modo Padrão HF) tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME) tokenizer.pad_token = tokenizer.eos_token model = AutoModelForCausalLM.from_pretrained( MODEL_NAME, load_in_4bit=True, # Quantização 4-bit padrão bitsandbytes device_map="auto", torch_dtype=torch.float16 ) # 3. Preparar LoRA sem depender do Unsloth model = prepare_model_for_kbit_training(model) peft_config = LoraConfig( r=16, lora_alpha=16, lora_dropout=0, target_modules=["q_proj", "k_proj", "v_proj", "o_proj"], task_type="CAUSAL_LM", ) model = get_peft_model(model, peft_config) # 4. Dataset dataset = load_dataset("json", data_files="seu_dataset.jsonl", split="train") # 5. SFTConfig (O jeito novo que evita o erro de 'tokenizer') sft_config = SFTConfig( output_dir="./peter-lora", max_seq_length=1024, # Reduzi de 2048 para não explodir a RAM do Space dataset_text_field="text", per_device_train_batch_size=1, gradient_accumulation_steps=4, learning_rate=2e-4, logging_steps=1, fp16=True, save_strategy="no", push_to_hub=False ) # 6. Trainer Corrigido trainer = SFTTrainer( model=model, train_dataset=dataset, args=sft_config, peft_config=peft_config, # Removido o argumento 'tokenizer' daqui para o SFTConfig assumir ) trainer.train() # 7. Salvar e Enviar model.push_to_hub("seuusername/peter-lora", private=True) tokenizer.push_to_hub("seuusername/peter-lora", private=True)