Reproducibility capsule

Regenerates the paper's data figures (Figs. 2-6) from IBM's public dataset LLMFineTuningBench.

Run

Needs Docker, running. In a terminal:

curl -L -o woais2.zip https://anonymous.4open.science/api/repo/woais2/zip
mkdir woais2 && cd woais2 && unzip -q ../woais2.zip
docker build -t woais2 .
docker run --rm -v ./out:/capsule/out woais2

The figures are in out/.

Without Docker: bash reproduce.sh (needs Python 3.11 and the font Times New Roman).

Checklist

1. Program: plot.py (Python 3.11, matplotlib 3.11.0, pandas 2.3.3).
2. Data: ibm-research/LLMFineTuningBench, 30,920 rows x 38 columns; if the Hub fails, the snapshot data/ado-sfttrainer.csv.
3. Environment: Docker 27.4, base image python:3.11.16-slim-trixie.
4. Hardware: none specific; tested on ARM64 (macOS) and x86-64.
5. Output: out/fig2_*.pdf to out/fig6_*.pdf, one PDF per figure of the paper (Figures 2-6).
6. Disk: about 1 GB (Docker image).
7. Time: about 2 minutes to build, under 1 minute to run.
8. Data license: Apache-2.0, IBM Research.