Datasetpublic

DariusTheGeek/mhqa-itu-artifacts

MHQA · ITU · Zindi Challenge — Artifacts DariusTheGeek/mhqa-itu-artifacts · the data + precomputed features that let the code repo reproduce submission sub_v40 (public LB 0.728509) for the ITU Multilingual Health QA in Low-Resource African Languages challenge. Code (which pulls this at runtime) lives on GitHub; trained weights are in the model repo DariusTheGeek/mhqa-itu-adapters. This is a reproducibility artifact bundle, not a raw dataset. It holds derived features and the… See the full description on the dataset page: https://huggingface.co/datasets/DariusTheGeek/mhqa-itu-artifacts.

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

0likes65downloads

Dataset Card

MHQA · ITU · Zindi Challenge — Artifacts

`DariusTheGeek/mhqa-itu-artifacts` · the data + precomputed features that let the code repo reproduce submission `sub_v40` (public LB 0.728509) for the ITU Multilingual Health QA in Low-Resource African Languages challenge. Code (which pulls this at runtime) lives on GitHub; trained weights are in the model repo `DariusTheGeek/mhqa-itu-adapters`.

This is a reproducibility artifact bundle, not a raw dataset. It holds derived features and the public Test inputs — not the labelled Train.csv / Val.csv, which are Zindi competition data (download from the competition page).

What's here

Files are stored at their repo-relative paths, so huggingface_hub.snapshot_download(local_dir=ROOT) drops each one exactly where the code expects it:

path	what
`data/Test.csv`, `data/SampleSubmission.csv`	public competition test inputs + submission format
`work/cache/*.parquet`	retrieval pools + precomputed selection features (incl. the frozen cross-encoder `ceft_*` outputs)
`work/submissions/sub_gen_v20.csv`	the generation base (medoid over the 3 adapters' K36 pools)
`work/submissions/sub_qlora_v2.csv`	reference generation used by the v30 selector
`work/cv/val_qlora_v2_preds.parquet`, `length_stats.json`	validation reference predictions
`artifacts/judge_reverts.json`	27 cached LLM-judge off-topic revert decisions (replayed offline)
`submissions/sub_v40.csv`	the reference scored submission (LB 0.728509)
`experiments/E1_gemma4_base/results/*.parquet`	gemma-4 reference predictions read by the selectors
`MANIFEST.sha256`	sha256 of every file — `sha256sum -c MANIFEST.sha256` proves byte-identity

How it's used

bash

# from the GitHub code repo:
bash stages/0_setup/fetch_artifacts.sh   # snapshot_download this repo -> work/ data/ artifacts/ …
bash reproduce.sh                         # rebuild sub_v40 byte-for-byte on CPU

Repo id and the pinned revision (v40) live in the code's configs/paths.yaml.

Provenance & licensing

All artifacts derive solely from the Zindi competition data (HASH consortium sexual/reproductive-health Q&A); no external or hand-authored data. Raw labelled Train.csv / Val.csv are not included (competition data). Use is governed by the Zindi competition rules; the accompanying models carry the Gemma / MedGemma licenses.

⚠️ Health content is for research reproduction only — not medical advice, not for clinical use.