docs: document per-task answer format and evaluation protocol

README.md

@@ -439,8 +439,8 @@ Same generators and parameters in both languages, so performance gaps isolate pr
 | `minesweeper` | Minesweeper with minimal hints preserving a unique solution | 100 / 100 / 100 | 100 / 100 / 100 | 600 |
 | `number_baseball` | Infer a hidden N-digit number from Strike/Ball hints | 100 / 100 / 100 | 100 / 100 / 100 | 600 |
 | `sat_puzzles` | Boolean satisfiability (CNF) framed as a natural-language scenario | 100 / 100 / 100 | 100 / 100 / 100 | 600 |
-| `sudoku` | 9x9 Sudoku with guaranteed-unique solutions | 100 / 100 / 100 | 100 / 100 / 100 | 600 |
-| `yacht_dice` | Assign 12 dice rolls to 12 scoring categories to maximize total score | 100 / 100 / 100 | 100 / 100 / 100 | 600 |
+| `sudoku` | 9x9 Sudoku with guaranteed-unique solutions; graded on the digits at 4-6 spot-checked cells, not the full grid | 100 / 100 / 100 | 100 / 100 / 100 | 600 |
+| `yacht_dice` | Assign 12 dice rolls to 12 scoring categories to maximize total score; graded on the partial sum over spot-checked rounds, not the 12-round total | 100 / 100 / 100 | 100 / 100 / 100 | 600 |
 
 ### 2. Script Adaptations (2 tasks · EN + KO)
 
@@ -468,15 +468,23 @@ No English counterpart — built on structures specific to the Korean language a
 ## Usage
 
 ```python
-from datasets import load_dataset
+from datasets import load_dataset, get_dataset_config_names
 
-# one task
-ds = load_dataset("HAERAE-HUB/NOLLI", "sudoku", split="test")
+# one task — `kinship` is one of the five Korean-only tasks
+ds = load_dataset("HAERAE-HUB/NOLLI", "kinship", split="test")
 
-# filter by language / difficulty (both are columns)
-ko_hard = ds.filter(lambda x: x["language"] == "ko" and x["difficulty"] == "hard")
+# language and difficulty are columns; the 10 bilingual tasks carry both en and ko
+cipher = load_dataset("HAERAE-HUB/NOLLI", "cipher", split="test")
+ko_hard = cipher.filter(lambda x: x["language"] == "ko" and x["difficulty"] == "hard")
+
+# all 15 tasks
+tasks = get_dataset_config_names("HAERAE-HUB/NOLLI")
+all_ds = {t: load_dataset("HAERAE-HUB/NOLLI", t, split="test") for t in tasks}
 ```
 
+Note that `kinship` carries an extra `choices` column, so the configs cannot be merged
+with `concatenate_datasets` without first aligning the schema.
+
 ## Structure
 
 Each config corresponds to one task. All configs share a single `test` split and the same schema
@@ -490,9 +498,22 @@ Each config corresponds to one task. All configs share a single `test` split and
 | `difficulty` | str | `easy` / `medium` / `hard` |
 | `question` | str | full prompt for the model |
 | `choices` | str | JSON-encoded answer choices — only present in the `kinship` config |
-| `answer` | str | gold answer |
+| `answer` | str | gold answer — format varies by task, see [Answer format](#answer-format) |
 | `solution` | str | step-by-step reference solution |
 
+## Answer format
+
+`answer` is not free text — the format is task-specific, and `question` already tells the model
+which one to produce. Two things to know before scoring.
+
+**Final line.** Seven tasks require a literal `Answer: <value>` last line — `inequality`, `jamo`,
+`korean_units`, `minesweeper`, `number_baseball`, `sudoku`, `yacht_dice`. The other eight take the
+bare value with no prefix.
+
+**Comparison.** Exact string match, with three exceptions: `sat_puzzles` (a JSON object) and
+`minesweeper` (a coordinate list) are order-insensitive, and `korean_units` carries a unit suffix
+that varies per item (`465373918평`).
+
 ## Citation
 
 ```bibtex