Xnhyacinth/Imagine-Data

Imagine

Question-answering tables for the Imagine / IAG-AAG research line, in one repository: the primary NQ, TQA, WQ views, the Natural-Questions context-count arms (ctxs1/ctxs5/ctxs100), and the original TQA splits (tqa-raw).

Subsets

Directory	Config / data_dir	Content	Purpose
NQ/	NQ	NQ question/answer, ctxs, compressed_ctxs_1..100	primary NQ view
TQA/	TQA	TQA question/answer/target, ctxs, compressed_ctxs_1..100	primary TQA view
WQ/	WQ	WQ question/answer, ctxs, compressed_ctxs_5..100	primary WQ view
ctxs1/	ctxs1	NQ with a single retrieved context	NQ context-count arm (k=1)
ctxs5/	ctxs5	NQ with 5 retrieved contexts	NQ context-count arm (k=5)
ctxs100/	ctxs100	NQ with 100 retrieved contexts	NQ context-count arm (k=100)
tqa-raw/	tqa-raw	original TQA splits as first published (the source repo's data/)	raw TQA table

Every subset has train, eval and test splits.

from datasets import load_dataset

# primary views
nq  = load_dataset("Xnhyacinth/Imagine-Data", "NQ",  split="test")
tqa = load_dataset("Xnhyacinth/Imagine-Data", "TQA", split="test")
wq  = load_dataset("Xnhyacinth/Imagine-Data", "WQ",  split="test")

# Natural-Questions context-count arms
c1   = load_dataset("Xnhyacinth/Imagine-Data", "ctxs1",   split="test")
c5   = load_dataset("Xnhyacinth/Imagine-Data", "ctxs5",   split="test")
c100 = load_dataset("Xnhyacinth/Imagine-Data", "ctxs100", split="test")

# original TQA splits
raw = load_dataset("Xnhyacinth/Imagine-Data", "tqa-raw", split="test")

The NQ/TQA/WQ views were previously also published as separate endpoints (Xnhyacinth/NQ-Image-Research-Data, Xnhyacinth/TQA-Image-Research-Data); both were merged into this repository and removed on 2026-10-13.

Merged subsets: provenance and verification

ctxs1/, ctxs5/ and ctxs100/ were merged from Xnhyacinth/Imagine-NQ-Contexts at revision c8428a340444086df850693ee704938dd230db2c; tqa-raw/ was merged from Xnhyacinth/Imagine-TQA at revision 3ea2dc07bba59c7d05a7e65f3d5cb1e8958b02ba. Files were copied server-side on the Hub and every one of the 57 parquet files was verified byte-identical to its source by size, git blob_id and LFS sha256 before the source repositories were removed. Paths are unchanged except data/ -> tqa-raw/, and the original source cards are reproduced below.

ARTIFACTS-ctxs-nq.json and ARTIFACTS-tqa-raw.json retain the merged subsets' own reviewed-copy records (source revision, whole-file before/after hashes, substitution counts).

Reviewed research copy

This complete file-set derivative of Xnhyacinth/Image at 863f15e0aaf57edada8ef3d3ff1b69a8427d426c applies reviewed substitutions to selected privacy-relevant string spans. Original configurations, splits, schemas, record IDs and all other fields are retained. Formal references, source attribution and documented examples remain subject to the original source terms; this card introduces no new license.

The ctxs1/ctxs5/ctxs100 subsets are a complete file-set derivative of Xnhyacinth/NQ-Image at c77aa8aaf052e9a779884faa34c60e86733da0fc with the same kind of reviewed substitutions (6 changed string fields across 6 files). The tqa-raw subset is the same kind of derivative of Xnhyacinth/TQA-Image at 60a12d654d748ea7ee6f7e52ce08bd2909e53bf8 (4 changed string fields).

Historical token counts, compression ratios, scores and other measured numeric fields are retained as source observations and have not been recomputed after text edits. The edits do not certify unchanged benchmark performance. Dataset byte metadata changes only by the edited strings' UTF-8 size delta and the changed files' stored-size delta.

Review covers native text and supported static metadata. It does not include OCR of image pixels or analysis of model tensor contents. No source code, models, inference or experiments were executed to produce this copy. Some redacted links require navigation from the retained formal source page.

ARTIFACTS.json records source revision, whole-file before/after hashes and substitution counts for the NQ/TQA/WQ views. It excludes literal private values, individual-value fingerprints, record offsets and review contexts. Original histories are retained in the restricted source repositories.

License

Contains Natural Questions data (upstream CC-BY-SA-3.0, attribution required). Per-source terms apply elsewhere; see the ARTIFACTS/README sections above.