| """Document processing for the MESA task tree. |
| |
| Two jobs: |
| |
| 1. Tier/domain routing. Each tier lives in its own dataset config, so a leaf |
| task loads only its tier and filters by domain here. The filters also drop |
| the contamination canary, which is the first record of every data file and |
| has no ``domain`` field. |
| |
| 2. Item rendering for continuation loglikelihood. MESA has two stem shapes and |
| they must not be rendered the same way: |
| |
| * **Cloze** stems carry a ``___`` blank mid-sentence (567 of the 727 English |
| items). The option belongs IN the blank, not after the stem. Appending |
| instead of substituting leaves the blank unfilled and makes every option |
| -- the key included -- ungrammatical, which is not a measurement of |
| anything. The context is the text before the blank and the continuation is |
| ``option + text after the blank``. |
| |
| * **Question** stems are complete questions. These get an ``Answer:`` cue |
| and the option is appended. |
| |
| The four options of a single item always render through the same branch, so |
| the ranking within an item is never comparing two different formats. |
| """ |
|
|
| import datasets |
|
|
| BLANK = "___" |
|
|
|
|
| |
|
|
| def _split_stem(stem): |
| """Return (before_blank, after_blank), or (None, None) for a question.""" |
| if BLANK not in stem: |
| return None, None |
| head, _, tail = stem.partition(BLANK) |
| return head.rstrip(), tail |
|
|
|
|
| def doc_to_text(doc): |
| head, _ = _split_stem(doc["stem"]) |
| if head is not None: |
| return head |
| return doc["stem"].rstrip() + "\nAnswer:" |
|
|
|
|
| def doc_to_choice(doc): |
| _, tail = _split_stem(doc["stem"]) |
| if tail is not None: |
| return [" " + c + tail for c in doc["choices"]] |
| return [" " + c for c in doc["choices"]] |
|
|
|
|
| |
|
|
| def _by_domain(domain): |
| def fn(dataset: datasets.Dataset) -> datasets.Dataset: |
| return dataset.filter(lambda d: d.get("domain") == domain) |
| return fn |
|
|
|
|
| filter_math = _by_domain("Math") |
| filter_english = _by_domain("English") |
| filter_science = _by_domain("Science") |
| filter_art = _by_domain("Art") |
|
|
