alea-institute/alea-doc-mc

ALEA flan2: doc_mc

Each row gives a real public-record passage and a four-option multiple-choice question whose wrong options are built from confusable content in the same passage (another party, date or amount, the negated or superseded value, the proposed rather than final value). About one in five questions asks which statement is NOT supported by the passage. Responses give the letter, the letter and option, or the letter with a one-sentence reason quoting the passage.

Known limitation: answerable from the options alone (not used in ALEA training)

After release, two blind options-only probes were run: a strong reviewer saw only the question and the four options, with no passage.

• 20 random items: 8 correct (40%).
• 60 items, 15 per answer letter (after the longest-option and single-minimal-pair items were removed from a revised build): 35 correct (58%). 20 of the 60 were answered through surface cues: an implausible or off-register distractor, overlap with the question, a single minimal pair, or the longest option. 8 more were answered from general knowledge.

Chance is 25%. The option-only check used during building (a weaker model guessing without the passage, rebalanced to about 27%) underestimated these cues. Many items can therefore be solved without reading the passage.

This dataset is not part of any ALEA training mix. A rebuilt version with distractors filtered against a strong options-only guesser is planned. Use these rows for multiple-choice reading only together with a passage-dependence check.

Sources

Passages (250-1,100 words) come from public-record KL3M data: 10-K annual reports (MD&A sections), state and federal court opinions (Caselaw Access Project), U.S. government web pages (guidance, news, FAQs), congressional hearing transcripts, material contracts filed as EDGAR exhibits, congressional committee reports, federal district court orders and opinions (RECAP), Federal Register notices, 8-K current reports, Federal Register proposed and final rules. No passage was edited, except for the deliberate twin variants described below, which are labelled as such in the gold record. Names of parties, judges, officials and companies are public record and are kept. Passages with private e-mail addresses or phone numbers were dropped. No more than 30% of documents come from one source pool, and no more than 5% from one court, agency, filer or host.

How the gold was made and verified

• Questions, the correct option, three distractors built from confusable content of the same passage, and evidence quotes were written by GLM-5.3-Flash.
• Code checks: evidence quotes occur verbatim in the passage, four distinct options, numbers in the correct option appear in its evidence, numbers in distractors come from the passage, no "all/none of the above".
• A second model from a different family (DeepSeek-V4.1-Flash) answered each item blind from the passage and the shuffled options; items were kept only when it chose the gold option.
• Guessability: on all items, the same model also answered from the question and options alone (no passage). It got 66% of the written items right, so items it got wrong are kept together with a random share of items it got right, per kind and source, such that its no-passage accuracy on the release is about 27% (chance 25%).
• Gold letters are balanced A-D by construction.

Rows that failed any check were dropped, never patched.

Size and composition

5,080 training rows from 4,204 source passages (5,312,392 tokens in the ALEA sw32k tokenizer). Presentation: plain instruction, Q/A frame, or chat turn (instruction or user turn always first).

kind	rows	share
mc_fact	4,131	81.3%
mc_not_supported	949	18.7%

source pool	rows	share
10k	1,017	20.0%
dotgov	716	14.1%
cap	714	14.1%
agr	710	14.0%
crpt	696	13.7%
chrg	675	13.3%
recap2	290	5.7%
fr	152	3.0%
8k	80	1.6%
frrule	30	0.6%

Checks against shortcuts

Bag-of-n-gram baselines that see only part of the input (cross-validated by source passage), compared with the majority-class prior:

{
 "mc_fact": {
  "options_only_lr": {
   "n": 4131,
   "acc": 0.275,
   "prior": 0.25,
   "lift": 0.025
  },
  "position_majority": {
   "acc": 0.2542,
   "prior": 0.25,
   "letters": {
    "A": 1050,
    "B": 1016,
    "D": 1046,
    "C": 1019
   }
  },
  "longest_option": {
   "acc": 0.2767,
   "prior": 0.25
  },
  "shortest_option": {
   "acc": 0.2452,
   "prior": 0.25
  },
  "llm_options_only_guess": {
   "n": 4131,
   "acc": 0.2694,
   "prior": 0.25
  }
 },
 "mc_not_supported": {
  "options_only_lr": {
   "n": 949,
   "acc": 0.2782,
   "prior": 0.25,
   "lift": 0.0282
  },
  "position_majority": {
   "acc": 0.2645,
   "prior": 0.25,
   "letters": {
    "B": 251,
    "C": 245,
    "D": 220,
    "A": 233
   }
  },
  "longest_option": {
   "acc": 0.2339,
   "prior": 0.25
  },
  "shortest_option": {
   "acc": 0.2034,
   "prior": 0.25
  },
  "llm_options_only_guess": {
   "n": 949,
   "acc": 0.2666,
   "prior": 0.25
  }
 },
 "all": {
  "options_only_lr": {
   "n": 5080,
   "acc": 0.2459,
   "prior": 0.25,
   "lift": -0.0041
  },
  "position_majority": {
   "acc": 0.2526,
   "prior": 0.25,
   "letters": {
    "A": 1283,
    "B": 1267,
    "D": 1266,
    "C": 1264
   }
  },
  "longest_option": {
   "acc": 0.2687,
   "prior": 0.25
  },
  "shortest_option": {
   "acc": 0.2374,
   "prior": 0.25
  },
  "llm_options_only_guess": {
   "n": 5080,
   "acc": 0.2689,
   "prior": 0.25
  }
 }
}

Decontamination and review

Rows sharing long n-grams with evaluation items or held-out sets were removed; overlaps made only of document boilerplate (headers and forms common to many source documents) were logged and kept. Held-out documents (about 1.5% of source documents) are not in this release. Builder spot checks read 45 rows against their sources. An independent blind review of 60 stratified rows found 1 wrong row (two options were supported). Items where one option contains another were removed afterwards.

Cost

LLM spend for this family: about $6.73.