llm-forensic-timeline

A work in progress repository for DFRWS APAC 2025 paper: A standardized methodology and dataset for evaluating LLM-based digital forensic timeline analysis.

The package covers the two halves of the methodology:

• Run — upload a log2timeline/Plaso CSV timeline to an LLM, replay the prompts of a task, and save every answer the model generates.
• Evaluate — score those answers against a manually validated ground truth with BLEU, ROUGE, and BERTScore, and write one CSV row per file.

Documentation: https://llm-forensic-timeline.readthedocs.io

Requirements

1. Create a virtual environment using venv (Python 3.10 or newer; BERTScore additionally needs 3.12+):

   python3.13 -m venv .venv

2. Activate the environment:

   source .venv/bin/activate (Windows: .venv\Scripts\activate)

3. Install the package:

   pip install -e ".[all]"

   Or, without cloning: pip install llm-forensic-timeline[all]

4. Open dataset repository on Zenodo

5. Download scenario-1.zip and unzip it

6. Copy all files from scenario-1/timeline/events directory to dataset/event-summarization directory

The extras split the optional halves: bertscore adds modern-bert-score, llm adds openai and python-dotenv, and all adds both. requirements.txt still holds the full environment used for the paper.

How to run

1. Add .env file in src directory. Add your OpenAI API keys in the .env file, something like: OPENAI_API_KEY=sk-proj-xxx

2. Run the task

   llm-forensic-timeline run --task event-summarization

   The original script still works: python src/run.py

How to evaluate

llm-forensic-timeline evaluate compares the LLM answers against the ground truth and writes the scores to a CSV file:

llm-forensic-timeline evaluate \
    --groundtruth-dir event-reconstruction/ground-truth-multiple \
    --llm-dir event-reconstruction/chatgpt-with-knowledge-multiple \
    --output event-reconstruction/chatgpt-with-knowledge-multiple-results.csv

The original script still works: edit the task and prediction variables at the top of src/evaluation-list.py, then run it from the directory that contains the task directory.

Three metrics are reported per file:

1. BLEU and ROUGE (ROUGE-1, ROUGE-2, ROUGE-L, ROUGE-Lsum) through the Hugging Face evaluate library. Both are n-gram overlap metrics, so they only measure lexical similarity.

2. BERTScore through modern-bert-score, which is semantic-aware and therefore credits an answer that is worded differently but carries the same meaning. This addresses the limitation of BLEU and ROUGE discussed in Sec. 4.4.2 of the paper. The columns bertscore_precision, bertscore_recall, and bertscore_f1 hold the mean over all entries of a file.

Notes on BERTScore:

• The model is LazerLambda/ModernBERT-large-ModBERTScore-19, the strongest checkpoint shipped with the package. It is the largest one, and its 8192-token context keeps long JSON entries of a forensic timeline intact, while the RoBERTa checkpoints truncate at 512 tokens. The weights (about 1.6 GB) are downloaded from Hugging Face on the first run and cached afterwards.
• Scores are rescaled with the baseline of the model, as recommended by the BERTScore authors. Without rescaling, scores of unrelated texts sit near 0.8; with rescaling, they sit near 0 and can be slightly negative, which makes the results easier to read.
• The script selects CUDA or Apple Silicon (MPS) automatically and falls back to the CPU. Evaluating a large timeline on the CPU is slow.
• Pass --no-bertscore to score with BLEU and ROUGE only, without loading a transformer model.

Development

Run the test suite:

pip install -e ".[test]"
pytest

Build the documentation locally:

pip install -r docs/requirements.txt
python -m sphinx -W -b html docs/source docs/_build/html

The Run Tests workflow runs pytest on Python 3.10 through 3.13 and builds the documentation on every push and pull request to main. Pushing a v*.*.* tag builds the distributions and publishes them to PyPI through OIDC Trusted Publishing, without any stored credentials.