Datasets:
docs: document per-task answer format and evaluation protocol
Documentation-only PR. No data files touched — the parquet files are untouched and the YAML frontmatter (15 configs, 7,500 examples) is unchanged.
While building an eval harness against this dataset I found the answer column hard to consume: the README describes it only as "gold answer", but the gold format differs substantially across the 15 tasks, and the output protocol the prompts mandate is not uniform either.
The main issue. 7 tasks instruct the model to end its response with a literal Answer: ... line (inequality, jamo, korean_units, minesweeper, number_baseball, sudoku, yacht_dice); the other 8 ask for the bare value with no prefix. A parser written against either convention alone silently drops half the benchmark, and nothing in the README says so.
Changes
New "Answer format" section:
- the final-line protocol split (which 7 tasks use
Answer:, which 8 do not); - a per-task gold-format table — e.g.
sat_puzzlesis a JSON name→bool object,minesweepera 0-based coordinate list,inequalitya flattened 5x5/6x6 grid,korean_unitsa number with its unit suffix where the unit varies per item (돈/푼/평/홉),kinshipa single A–Z letter,timeone of three shapes (YYYY.M.D/ 60갑자 일진 / 요일); - a reference exact-match scorer. It round-trips 7500/7500 against the gold answers in this repo and rejects wrong values, so it should be a safe starting point for anyone evaluating.
- the final-line protocol split (which 7 tasks use
Two task descriptions corrected against the actual data:
sudoku— graded on the digits at 4–6 spot-checked cells, not the full grid;yacht_dice— graded on a partial sum over spot-checked rounds (2 for easy/medium, 8 for hard), not the 12-round total. This is 600/600 of the config.
Usage section: note that no default config is declared (a bare
load_dataset("HAERAE-HUB/NOLLI")errors), show how to load all 15 configs, and flag thatkinship's extrachoicescolumn blocks a naiveconcatenate_datasets.
Also worth noting: I verified the data itself and it is clean — 7,500 rows, exactly 100 per task x language x difficulty cell, zero duplicate ids, zero empty question/answer/solution fields, and the YAML num_examples match the parquet contents exactly.
Happy to adjust wording or drop any part of this.
Revised down from +89 to +31/-10 after review — the first draft was too long for a dataset card.
Kept
- The Usage rewrite: a
get_dataset_config_namesloop for loading all 15 configs, and a note thatkinship's extrachoicescolumn blocks a naiveconcatenate_datasets. The single-task example now useskinship(Korean-only) and the language filter usescipher, since filtering a Korean-only config bylanguage == "ko"is a no-op. - The
sudokuandyacht_dicerow corrections, now stated fully in the task table itself (not the full grid,not the 12-round total) rather than repeated in prose. - An
Answer formatsection, cut to the two things that actually trip people up: which 7 tasks require a literalAnswer:last line vs the other 8, and the three tasks that are not plain string comparisons (sat_puzzles,minesweeper,korean_units).
Dropped
- The 40-line reference scorer — it belongs in the GitHub repo, not the card. I have a version that round-trips 7500/7500 against the gold answers here if you want it there.
- The per-task gold-format table (15 rows). Accurate, but it duplicates what one
print(ds[0])shows and would need maintenance on every generator change. - "There is no default config" — the frontmatter lists 15
configs:, so the line reads as a contradiction, andload_datasetwithout a config already errors with all 15 names listed.
Supersedes #2, which I am closing — the yacht_dice line is included here.
Note: I could not execute the Usage snippet end-to-end in my environment (datasets 3.6.0 crashes on Python 3.14 via dill). The facts it encodes are verified against the parquet files directly — kinship has choices, cipher ko+hard is 100 rows, 15 configs / 7,500 rows total — but the snippet itself is unrun, so a quick check before merging is worthwhile.