`MAD_human_labelled_dataset.json` spans three taxonomy versions — is this documented anywhere?

#3
by hknsahin97 - opened

Thanks for keeping this dataset current — the 2026-08-13 "Rebuild trace-label join" update clearly did real work, and it's the reason I'm posting here rather than only on GitHub.

While computing inter-annotator agreement on MAD_human_labelled_dataset.json I found that its 19 records are not annotated in one taxonomy. Grouping records by the set of mode codes in their annotations array gives three distinct sets:

round records modes
Round 1 5 18
Round 2, Round 3 10 17
Generlazability 4 14 — the published MAST

Mode descriptions were carried over verbatim between versions, so the versions can be aligned by description text. Doing that shows the code numbers were reassigned: only 2 of 14 codes are unchanged between Round 1 and MAST-14, and 7 of 14 between Rounds 2–3 and MAST-14. For example 2.2 is Step Repetition in Round 1 and Fail to Ask for Clarification in MAST-14.

Two things follow, and I'd like to check both with you rather than assume:

  1. The dataset card documents one schema. It describes trace as {key, index, trajectory} and the label field as mast_annotation. That holds for MAD_full_dataset.json. In the human-labelled file trace is a plain string and the label field is annotations with a different, longer mode list. A reader following the card will mis-parse the second file.

  2. Code-level comparison against MAST-14 is invalid for 15 of the 19 records unless remapped by round. This looks like a sufficient explanation for the κ = 0.05 reported in MAST issue #13, without needing the "different trace set" hypothesis raised there — and it is falsifiable: a round-aware remapping should raise that κ substantially. If it doesn't, my alignment is wrong.

Separately, and reproducing on the current file: the 19 records carry only 8 distinct annotations blocks, with 18 of 19 sharing a byte-identical block with another record; every duplicate group is within a single round and spans multiple frameworks.

Full detail, the version mapping table, and a script that reproduces every number (stdlib only, downloads the files itself):

https://github.com/hknsahin97-coder/mad-audit

I also opened MAST issue #18 with the same findings, but the GitHub repo hasn't been pushed to since 2025-07, whereas this dataset is clearly maintained — so this is probably the better place to reach you.

If the version mixing is deliberate — a record of how the taxonomy developed rather than a packaging accident — a sentence in the card and a published cross-version code mapping would close it entirely. Happy to be told I'm misreading something.

Sign up or log in to comment