SPDX-FileContributor	Arthit Suriyawongkul
SPDX-FileCopyrightText	2024-present Arthit Suriyawongkul <suriyawa@tcd.ie>
SPDX-FileType	DOCUMENTATION
SPDX-License-Identifier	CC0-1.0

Sentiment Demo: A Simple AI Application and its AI BOM Example

A simple text classification application, published solely to demonstrate a software bill of materials (SBOM) in SPDX 3.0 format.

The main content of the package are software bill of materials at bom.spdx3.json (written by hand) and bom-generated.spdx3.json (Pitloom-generated). Other files are given just to complete the illustration.

Not recommended for actual text classification tasks.

Updates:

• July 2026: Version 0.3.1 - SBOM fragments that were generated during the runs of evaluate/predict/preprocess/train.py are merged into the main SBOM -- shipped with the Python wheel. This gives more data/model provenance.
• March 2026: Adopted as a development reference for the Pitloom SBOM generator.
• May 2025: Added to the SPDX Usage Examples repository as ai/example02.

SBOM demonstration design goals:

• Comprehensible: Small enough for a human to understand easily.
• Informative: Elaborate enough to showcase the use of various information fields within an SBOM.
• Testable: Designed to facilitate testing and evaluation against specific use case requirements.

For more information about implementing AI BOM using SPDX specification, see Karen Bennet, Gopi Krishnan Rajbahadur, Arthit Suriyawongkul, and Kate Stewart, “Implementing AI Bill of Materials (AI BOM) with SPDX 3.0: A Comprehensive Guide to Creating AI and Dataset Bill of Materials”, The Linux Foundation, October 2024.

Content

.
├── LICENSE                    License information
├── README.md                  This README file
├── bom-generated.spdx3.json   SBOM in SPDX 3 format (generated)
├── bom.spdx3.json             SBOM in SPDX 3 format (handwritten)
├── data                       Dataset, preprocessed and tokenized
│   ├── test.txt               Testing data
│   ├── train.txt              Training data
│   └── valid.txt              Validation data
├── rawdata                    Raw dataset, before preprocessing
│   ├── test                   Testing data
│   │   ├── neg.txt            Testing samples for label "neg" (negative)
│   │   ├── neu.txt            Testing samples for label "neu" (neutral)
│   │   ├── pos.txt            Testing samples for label "pos" (positive)
│   │   └── q.txt              Testing samples for label "q" (question)
│   ├── train                  Training data
│   │   └── ...
│   └── valid                  Validation data
│       └── ...
├── src
│   ├── evaluate.py            Script to evaluate prediction performance
│   ├── model.bin              Sentiment analysis model
│   ├── predict.py             Script to predict a label of a text
│   ├── preprocess.py          Script to prepare training data
│   └── train.py               Script to build a model
└── techdocs                   Technical documentation
    ├── dataprepare.md         Data preparation
    └── instructions.md        Instruction for use

Usage

See instruction for use for how to use the application.

Data preparation

See data preparation.

Generating the SBOM

Run the following commands to generate SBOM fragments at each stage of the AI development lifecycle. Finally, use Pitloom to merge these fragments with the core project SBOM into a single, comprehensive file.

SBOM fragment 1: Preprocess the data

python src/sentimentdemo/preprocess.py rawdata data

This processes the raw data into the data/ directory and generates an SBOM fragment at fragments/preprocess.spdx3.json (as specified by loom.run in preprocess.py).

SBOM fragment 2: Train the model

AUTOTUNE_DURATION=5 python src/sentimentdemo/train.py data/train.txt data/valid.txt src/sentimentdemo/model.bin

This step outputs the training fragment to fragments/train.spdx3.json.

Note: We use AUTOTUNE_DURATION=5 to speed up the training process. Since the goal of this demo is to illustrate SBOM generation rather than produce a highly performant model, a short duration is sufficient.

SBOM fragment 3: Run predictions

python src/sentimentdemo/predict.py "test"

This step outputs the prediction fragment to fragments/predict.spdx3.json.

Production note: Generating a fragment on every prediction may introduce unnecessary overhead. You should typically enable this only during testing and packaging, and disable it before deploying to production.

SBOM fragment 4: Evaluate model

python src/sentimentdemo/evaluate.py src/sentimentdemo/model.bin data/test.txt

This step outputs the evaluation fragment to fragments/evaluate.spdx3.json.

Merge and generate the final SBOM

pitloom project . --pretty --output bom-generated.spdx3.json

Pitloom collects all the generated fragments in the fragments/ directory and merges them. The final, consolidated SBOM is saved to bom-generated.spdx3.json in your current working directory.

Notes

• Development is in the main branch.
• The diagram is generated from a PlantUML file: bom.spdx.puml. The PlantUML file is generated by spdx3ToGraph. To brevity, spdxIds and long strings are shortened by the shortenid.sh script in tools/, and all but one hyperparameter have been manually removed.
• The energy used by the computer during model training is tracked by energy-tracker. It measures how much energy the computer uses during the training. This means the actual energy used for training the model might be a bit less than the reported amount.
• Both handwritten SBOM and generated SBOM are validated by spdx3-validate, which use the JSON Schema from https://spdx.org/schema/3.0.1/spdx-json-schema.json and the SHACL from https://spdx.org/rdf/3.0.1/spdx-model.ttl.
• Next steps:
  • Add external dependency relationships (e.g. dependsOn, hasProvidedDependency)
  • Using information requirements and obligations in the EU AI Act as a target, labeling all relevant properties and relationships with corresponding difficulty levels and support levels, based on the BOM Maturity Model.

Licenses

Apart from the data and components listed in the table below, the code and content in this repository are dedicated to the public domain under the terms of Creative Commons Zero ("CC0") 1.0 Universal, which have no copyright and related or neighboring rights worldwide to the extent allowed by law.

Component	Name	License	Notes
Training data	Wisesight Sentiment Corpus	CC0-1.0	Samples from the corpus are in rawdata/. Preprocessed data is in data/. See data preparation for details.
Text preprocessor	th-simple-preprocessor	Apache-2.0
Word tokenizer	newmm-tokenizer	Apache-2.0	Inherited the license from PyThaiNLP.
Text classifier	fastText	MIT	Use fasttext-community, which is a community-maintained fork.
Array package	NumPy	BSD-3-Clause AND 0BSD AND MIT AND Zlib AND CC0-1.0

The specific version information can be found in pyproject.toml.

Citation

If you use this software, including its software bill of materials (SBOM), please cite it as follows:

Suriyawongkul, Arthit. “Sentiment Demo: A Simple AI Application and Its AI BOM Example”. Zenodo, 8 November 2024. https://doi.org/10.5281/zenodo.14055332.

BibTeX:

@software{Suriyawongkul_Sentiment_Demo_A_2024,
    author = {Suriyawongkul, Arthit},
    doi = {10.5281/zenodo.14055332},
    license = {CC0-1.0},
    month = nov,
    title = {{Sentiment Demo: A Simple AI Application and its AI BOM Example}},
    url = {https://github.com/bact/sentimentdemo/},
    version = {0.1},
    year = {2024}
}