Labspace - Fine tuning

This Labspace provides a hands-on walkthrough on fine-tuning models using Docker Offload, Docker Model Runner, and Unsloth.

Learning objectives

This Labspace will teach you the following:

• Use Docker Offload to fine tune a model
• Package and share the model on Docker Hub
• Run the custom model with Docker Model Runner

Launch the Labspace

To launch the Labspace, run the following command:

docker compose -f oci://dockersamples/labspace-fine-tuning up -d

And then open your browser to http://localhost:3030.

Using the Docker Desktop extension

If you have the Labspace extension installed (docker extension install dockersamples/labspace-extension if not), you can also click this link to launch the Labspace.

Acknowledgements

This repository contains an example of fine-tuning a language model for PII (Personally Identifiable Information) masking using the Unsloth framework.

Special thanks to AI4Privacy for providing the comprehensive PII masking dataset that makes this fine-tuning example possible.

Dataset Source

The training dataset data/training_data.json used in this example was created from the AI4Privacy PII Masking 400k Dataset:

• Original Dataset: ai4privacy/pii-masking-400k
• Dataset Description: World's largest open dataset for privacy masking with 406,896 entries
• Languages Supported: English, Italian, French, German, Dutch, Spanish
• PII Classes: 17 different types of PII in the public dataset

Licensing and Attribution

Important License Notice

The original dataset is provided by AI4Privacy under a custom license with the following requirements:

• Academic Use: Encouraged with proper citation
• Commercial Use: Requires licensing from AI4Privacy
  • Contact: [email protected]
• Attribution: This project uses data derived from the ai4privacy/pii-masking-400k dataset

Citation

If you use this dataset or derivatives in academic work, please cite:

@dataset{ai4privacy_pii_masking_400k,
  title={PII Masking 400k Dataset},
  author={AI4Privacy},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/ai4privacy/pii-masking-400k}
}

Data Preparation Process

The dataset was prepared using the prepare_pii_masking_for_unsloth.py script, which:

1. Loads the original dataset from HuggingFace: ai4privacy/pii-masking-400k
2. Transforms the data into a format suitable for fine-tuning with Unsloth
3. Creates instruction-following pairs with two modes:
   • Redaction mode: Maps source text to masked text with PII labels
   • Spans mode: Extracts PII spans with their positions and labels

Using the Data Preparation Script

# Basic usage - creates redaction mode dataset
python prepare_pii_masking_for_unsloth.py --outdir data_out

# Filter by languages (e.g., English and Spanish)
python prepare_pii_masking_for_unsloth.py --langs en es --outdir data_out_en_es

# Filter by locales (e.g., US and GB)
python prepare_pii_masking_for_unsloth.py --locales US GB --outdir data_out_us_gb

# Use spans mode instead of redaction
python prepare_pii_masking_for_unsloth.py --mode spans --outdir spans_out

# Add custom EOS token
python prepare_pii_masking_for_unsloth.py --outdir data_out --eos "</s>"

# Sample a subset for testing
python prepare_pii_masking_for_unsloth.py --outdir data_out --sample_train 1000 --sample_val 200

Fine-Tuning Process

The finetune.py script demonstrates how to:

1. Load a pre-trained model (Gemma-3-270M-IT)
2. Prepare the dataset for instruction fine-tuning
3. Configure LoRA adapters for efficient training
4. Train the model using the SFTTrainer from TRL
5. Save the fine-tuned model