HuggingFace Hub publishing

from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments
from trl import SFTTrainer, SFTConfig
from peft import LoraConfig
import torch
from datasets import Dataset

# Login to HuggingFace Hub first (run once)
# from huggingface_hub import login
# login(token="hf_your_token_here")

# Create a minimal training dataset
dataset_dict = {
    "text": [
        "The quick brown fox jumps over the lazy dog. This is a fine-tuning example.",
        "Machine learning models require careful tuning and validation steps.",
        "HuggingFace transformers make it easy to fine-tune large language models.",
    ]
}
dataset = Dataset.from_dict(dataset_dict)

# Configure model and training
model_name = "gpt2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Add padding token for GPT-2
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# LoRA config for efficient fine-tuning
lora_config = LoraConfig(
    r=8,
    lora_alpha=32,
    lora_dropout=0.1,
    target_modules=["c_attn", "c_proj"],
    bias="none",
    task_type="CAUSAL_LM"
)

# Training config WITH hub publishing
sft_config = SFTConfig(
    output_dir="./fine_tuned_gpt2",
    num_train_epochs=2,
    per_device_train_batch_size=2,
    max_seq_length=128,
    learning_rate=2e-4,
    # Hub publishing settings
    push_to_hub=True,
    hub_model_id="neural-base-demo/gpt2-finetuned",
    hub_strategy="every_save",
    hub_private_repo=False,
    save_strategy="epoch",
    save_total_limit=2,
)

# Create trainer
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    args=sft_config,
    train_dataset=dataset,
    peft_config=lora_config,
    dataset_text_field="text",
)

# Train and automatically push to Hub on each save
trainer.train()

# Verify it was pushed
print("Model pushed to Hub at: https://huggingface.co/neural-base-demo/gpt2-finetuned")

# Load your own model from Hub to verify
from transformers import AutoModel
downloaded_model = AutoModel.from_pretrained("neural-base-demo/gpt2-finetuned", trust_remote_code=True)
print("Successfully loaded model from Hub")