Train a Llama model from scratch

In this tutorial, we'll walk through the process of training a language model using the Llama model architecture and the Transformers library.

1. Installing the Required Libraries

We'll start by installing the necessary libraries using pip:

!pip install -q datasets accelerate evaluate trl accelerate transformers jinja2

2. Logging into Hugging Face Hub

Next, we'll log into the Hugging Face Hub to access the required models and datasets:

from huggingface_hub import notebook_login

notebook_login()

3. Loading the Necessary Libraries and Models

We'll import the required libraries and load the Llama model and tokenizer:

this part is pretty complicated, so stay with me.

from datasets import load_dataset

dataset = load_dataset("your_dataset_name", split="train") # load the dataset

Here, we'll get the corpus to past to the tokenizer

def get_training_corpus():
    for i in range(0, len(dataset), 1000):
        yield dataset[i : i + 1000]["text"]

training_corpus = get_training_corpus()

The base tokenizer is up to you, I'm using a blank one, but a lot of people opt for different ones, such as gpt2.

from tokenizers import ByteLevelBPETokenizer

tokenizer = ByteLevelBPETokenizer()
tokenizer.train_from_iterator(
    training_corpus,
    vocab_size=3200,
    min_frequency=2,
    special_tokens=["<s>", "<pad>", "</s>", "<unk>", "<mask>", "<|user|>", "<|bot|>", "<|end|>"] # you can pick the last two or three, as you'll see next
)

Next, we'll define the tokenizer special tokens and chat template.

from transformers import PreTrainedTokenizerFast

special_tokens = {
    "bos_token": "<s>",
    "eos_token": "</s>",
    "unk_token": "<unk>",
    "pad_token": "<pad>",
    "mask_token": "<mask>",
    "additional_special_tokens": ["<|user|>", "<|bot|>", "<|end|>"] # same here
}
tokenizer.add_special_tokens(special_tokens)

tokenizer.user_token_id = tokenizer.convert_tokens_to_ids("<|user|>") # here
tokenizer.assistant_token_id = tokenizer.convert_tokens_to_ids("<|bot|>") # too

chat_template = "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '<|user|>\n' + message['content'] + '<|end|>\n' }}{% elif message['role'] == 'assistant' %}{{ '<|bot|>\n' + message['content'] + '<|end|>\n' }}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}{{ eos_token }}" # this is where you define the chat template, so you can go crazy here. Something a lot of people do for whatever reason is add seamingly random newline characters

tokenizer.chat_template = chat_template

Now, finally, we'll define the model.

from transformers import LlamaConfig, LlamaForCausalLM

print(tokenizer.apply_chat_template([{"role": "user", "content": "Why is the sky blue?"}, {"role": "assistant", "content": "Due to rayleigh scattering."}], tokenize=False)) # test to see if the chat template worked

config = LlamaConfig(
    vocab_size=tokenizer.vocab_size,
    hidden_size=512,
    intermediate_size=1024,
    num_hidden_layers=8,
    num_attention_heads=8,
    max_position_embeddings=512,
    rms_norm_eps=1e-6,
    initializer_range=0.02,
    use_cache=True,
    pad_token_id=tokenizer.pad_token_id,
    bos_token_id=tokenizer.bos_token_id,
    eos_token_id=tokenizer.eos_token_id,
    tie_word_embeddings=False,
)

model = LlamaForCausalLM(config)

4. Formatting the Dataset

We'll define a function to format the prompts in the dataset and map the dataset:

def format_prompts(examples):
    """
    Define the format for your dataset
    This function should return a dictionary with a 'text' key containing the formatted prompts.
    """
    pass

dataset = dataset.map(format_prompts, batched=True)

print(dataset['text'][2]) # Check to see if the fields were formatted correctly

5. Setting Up the Training Arguments

Define the training args:

from transformers import TrainingArguments

args = TrainingArguments(
    output_dir="your_output_dir",
    num_train_epochs=4, # replace this, depending on your dataset
    per_device_train_batch_size=16,
    learning_rate=1e-4,
    optim="sgd" # sgd, my beloved
)

6. Creating the Trainer

We'll create an instance of the SFTTrainer from the trl library:

from trl import SFTTrainer

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    args=args,
    train_dataset=dataset,
    dataset_text_field='text',
    max_seq_length=512
)

7. Training the Model

Finally, we'll start the training process:

trainer.train()

8. Pushing the Trained Model to Hugging Face Hub

After the training is complete, you can push the trained model to the Hugging Face Hub using the following command:

trainer.push_to_hub()

This will upload the model to your Hugging Face Hub account, making it available for future use or sharing.

That's it!