Fine-tune a SmolLM on domain-specific synthetic data from a LLM

Yes, smoll models can beat GPT4-like models on domain-specific tasks but don't expect miracles. When comparing smoll vs large, consider all costs and gains like difference performance and the value of using private and local models and data that you own and can do with as you please.

The Hugging Face SmolLM models are blazingly fast and remarkably powerful. With its 135M, 360M and 1.7B parameter models, it is a great choice for a small and fast model. The great thing about SmolLM is that it is a general-purpose model that can be fine-tuned on domain-specific data.

A lack of domain-specific datasets is a common problem for smaller and more specialized models. This is because it is difficult to find a dataset that is both representative and diverse enough for a specific task. We solve this problem by generating a synthetic dataset from an LLM using the synthetic-data-generator, which is available as a Hugging Face Space or on GitHub.

In this example, we will fine-tune a SmolLM2 model on a synthetic dataset generated from meta-llama/Meta-Llama-3.1-8B-Instruct with the synthetic-data-generator.

Install the dependencies

We will install some basic dependencies for the fine-tuning with trl but we will use the Synthetic Data Generator UI to generate the synthetic dataset.

!pip install transformers datasets trl torch

The problem

Reasoning data has proven to be a fundamental change in the performance of generative models. Reasoning is amazing but it also means the model generates more "chatty" during the token generation process, causing the model to become slower and more expensive. For this reason, we want to create a model that can reason without being too chatty. Therefore, we will generate a concise reasoning dataset and fine-tune a SmolLM2 model on it.

Let's generate some data

Let's go to the hosted Hugging Face Space to generate the data. This is done in three steps 1) we come up with a dataset description, 2) iterate on the task configuration, and 3) generate and push the data to Hugging Face. A more detailed flow can be found in this blog post.

For this example, we will generate 5000 chat data examples for a single turn in the conversation. All examples have been generated with a temperature of 1. After some iteration, we come up with the following system prompt:

You are an AI assistant who provides brief and to-the-point responses with logical step-by-step reasoning. Your purpose is to offer straightforward explanations and answers so that you can get to the heart of the issue. Respond with extremely concise, direct justifications and evidence-based conclusions. User questions are direct and concise.

We press the "Push to Hub" button and wait for the data to be generated. This takes a few hours and we end up with a dataset with 5000 examples, which is the maximum number of examples we can generate in a single run. You can scale this by deploying a private instance of the Synthetic Data Generator.

The data is pushed to Argilla too so we recommend inspecting and validating the the data before finetuning the actual model. We applied some basic filters and transformations to the data to make it more suitable for fine-tuning.

Fine-tune the model

We will use TRL to fine-tune the model. It is part of the Hugging Face ecosystem and works seamlessly on top of datasets generated by the synthetic data generator without needing to do any data transformations.

Load the model

We will first load the model and tokenizer and set up the chat format.

# Import necessary libraries
from transformers import AutoModelForCausalLM, AutoTokenizer
from datasets import load_dataset
from trl import SFTConfig, SFTTrainer, setup_chat_format
import torch
import os

device = (
    "cuda"
    if torch.cuda.is_available()
    else "mps" if torch.backends.mps.is_available() else "cpu"
)

# Load the model and tokenizer
model_name = "HuggingFaceTB/SmolLM2-360M"
model = AutoModelForCausalLM.from_pretrained(
    pretrained_model_name_or_path=model_name
)
tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=model_name)

# Set up the chat format
model, tokenizer = setup_chat_format(model=model, tokenizer=tokenizer)

Test the base model

We will first test the base model to see how it performs on the task. During this step we will also generate a prompt for the model to respond to, to see how it performs on the task.

from transformers import pipeline
# Let's test the base model before training
prompt = "What is the primary function of mitochondria within a cell?"

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device=device)
pipe(prompt, max_new_tokens=100)
# [{'generated_text': 'What is the primary function of mitochondria within a cell?\n\The function of the mitochondria is to produce energy for the cell through a process called cellular respiration.'}]

Load the dataset

For fine-tuning, we need to load the dataset and tokenize it. We will use the synthetic-concise-reasoning-sft-filtered dataset that we generated in the previous step.

from datasets import load_dataset

ds = load_dataset("argilla/synthetic-concise-reasoning-sft-filtered")
def tokenize_function(examples):
    examples["text"] = tokenizer.apply_chat_template([{"role": "user", "content": examples["prompt"].strip()}, {"role": "assistant", "content": examples["completion"].strip()}], tokenize=False)
    return examples
ds = ds.map(tokenize_function)

Fine-tune the model

We will now fine-tune the model. We will use the SFTTrainer from the trl library to fine-tune the model. We will use a batch size of 4 and a learning rate of 5e-5. We will also use the use_mps_device flag to use the MPS device if available.

os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"

# Configure the SFTTrainer
sft_config = SFTConfig(
    output_dir="./sft_output",
    num_train_epochs=1,
    per_device_train_batch_size=4,  # Set according to your GPU memory capacity
    learning_rate=5e-5,  # Common starting point for fine-tuning
    logging_steps=100,  # Frequency of logging training metrics
    use_mps_device= True if device == "mps" else False,
    hub_model_id="argilla/SmolLM2-360M-synthetic-concise-reasoning",  # Set a unique name for your model
    push_to_hub=True,
)

# Initialize the SFTTrainer
trainer = SFTTrainer(
    model=model,
    args=sft_config,
    train_dataset=ds["train"],
    tokenizer=tokenizer,
)
trainer.train()
# {'loss': 1.4498, 'grad_norm': 2.3919131755828857, 'learning_rate': 4e-05, 'epoch': 0.1}
# {'loss': 1.362, 'grad_norm': 1.6650595664978027, 'learning_rate': 3e-05, 'epoch': 0.19}
# {'loss': 1.3778, 'grad_norm': 1.4778285026550293, 'learning_rate': 2e-05, 'epoch': 0.29}
# {'loss': 1.3735, 'grad_norm': 2.1424977779388428, 'learning_rate': 1e-05, 'epoch': 0.39}
# {'loss': 1.3512, 'grad_norm': 2.3498542308807373, 'learning_rate': 0.0, 'epoch': 0.48}
# {'train_runtime': 1911.514, 'train_samples_per_second': 1.046, 'train_steps_per_second': 0.262, 'train_loss': 1.3828572998046875, 'epoch': 0.48}

For the example, we did not use a specific validation set but we can see the loss is decreasing, so we assume the model is generalising well to the training data. To get a better understanding of the model's performance, let's test it again with the same prompt.

Run inference

We can now run inference with the fine-tuned model.

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
pipe(prompt, max_new_tokens=100)
# [{'generated_text': 'The primary function of mitochondria is to generate energy for the cell. They are organelles found in eukaryotic cells that convert nutrients into ATP (adenosine triphosphate), which is the primary source of energy for cellular processes.'}]

Conclusion

We have fine-tuned a SmolLM2 model on a synthetic dataset generated from a large language model. We have seen that the model performs well on the task and that the synthetic data is a great way to generate diverse and representative data for supervised fine-tuning.

In practice, you would likely want to spend more time on the data quality and fine-tuning the model but the flow shows the Synthetic Data Generator is a great tool to generate synthetic data for any task.

Overall, I think it is pretty cool for a couple of hours of generation and fine-tuning on consumer hardware.