README.md

metadata

license: apache-2.0
pretty_name: ProofWriter (The Cut)
tags:
  - logical-reasoning
  - proof-generation
configs:
  - config_name: owa-att-noneg-d1
    data_files:
      - split: train
        path: owa-att-noneg-d1/train.parquet
      - split: val
        path: owa-att-noneg-d1/val.parquet
      - split: test
        path: owa-att-noneg-d1/test.parquet
  - config_name: owa-att-noneg-d2
    data_files:
      - split: train
        path: owa-att-noneg-d2/train.parquet
      - split: val
        path: owa-att-noneg-d2/val.parquet
      - split: test
        path: owa-att-noneg-d2/test.parquet
  - config_name: owa-att-noneg-d3
    data_files:
      - split: train
        path: owa-att-noneg-d3/train.parquet
      - split: val
        path: owa-att-noneg-d3/val.parquet
      - split: test
        path: owa-att-noneg-d3/test.parquet
  - config_name: owa-att-noneg-d5
    data_files:
      - split: train
        path: owa-att-noneg-d5/train.parquet
      - split: val
        path: owa-att-noneg-d5/val.parquet
      - split: test
        path: owa-att-noneg-d5/test.parquet
  - config_name: owa-att-neg-d1
    data_files:
      - split: train
        path: owa-att-neg-d1/train.parquet
      - split: val
        path: owa-att-neg-d1/val.parquet
      - split: test
        path: owa-att-neg-d1/test.parquet
  - config_name: owa-att-neg-d2
    data_files:
      - split: train
        path: owa-att-neg-d2/train.parquet
      - split: val
        path: owa-att-neg-d2/val.parquet
      - split: test
        path: owa-att-neg-d2/test.parquet
  - config_name: owa-att-neg-d3
    data_files:
      - split: train
        path: owa-att-neg-d3/train.parquet
      - split: val
        path: owa-att-neg-d3/val.parquet
      - split: test
        path: owa-att-neg-d3/test.parquet
  - config_name: owa-att-neg-d5
    data_files:
      - split: train
        path: owa-att-neg-d5/train.parquet
      - split: val
        path: owa-att-neg-d5/val.parquet
      - split: test
        path: owa-att-neg-d5/test.parquet
  - config_name: owa-rel-noneg-d1
    data_files:
      - split: train
        path: owa-rel-noneg-d1/train.parquet
      - split: val
        path: owa-rel-noneg-d1/val.parquet
      - split: test
        path: owa-rel-noneg-d1/test.parquet
  - config_name: owa-rel-noneg-d2
    data_files:
      - split: train
        path: owa-rel-noneg-d2/train.parquet
      - split: val
        path: owa-rel-noneg-d2/val.parquet
      - split: test
        path: owa-rel-noneg-d2/test.parquet
  - config_name: owa-rel-noneg-d3
    data_files:
      - split: train
        path: owa-rel-noneg-d3/train.parquet
      - split: val
        path: owa-rel-noneg-d3/val.parquet
      - split: test
        path: owa-rel-noneg-d3/test.parquet
  - config_name: owa-rel-noneg-d5
    data_files:
      - split: train
        path: owa-rel-noneg-d5/train.parquet
      - split: val
        path: owa-rel-noneg-d5/val.parquet
      - split: test
        path: owa-rel-noneg-d5/test.parquet
  - config_name: owa-rel-neg-d1
    data_files:
      - split: train
        path: owa-rel-neg-d1/train.parquet
      - split: val
        path: owa-rel-neg-d1/val.parquet
      - split: test
        path: owa-rel-neg-d1/test.parquet
  - config_name: owa-rel-neg-d2
    data_files:
      - split: train
        path: owa-rel-neg-d2/train.parquet
      - split: val
        path: owa-rel-neg-d2/val.parquet
      - split: test
        path: owa-rel-neg-d2/test.parquet
  - config_name: owa-rel-neg-d3
    data_files:
      - split: train
        path: owa-rel-neg-d3/train.parquet
      - split: val
        path: owa-rel-neg-d3/val.parquet
      - split: test
        path: owa-rel-neg-d3/test.parquet
  - config_name: owa-rel-neg-d5
    data_files:
      - split: train
        path: owa-rel-neg-d5/train.parquet
      - split: val
        path: owa-rel-neg-d5/val.parquet
      - split: test
        path: owa-rel-neg-d5/test.parquet
  - config_name: cwa-att-noneg-d1
    data_files:
      - split: train
        path: cwa-att-noneg-d1/train.parquet
      - split: val
        path: cwa-att-noneg-d1/val.parquet
      - split: test
        path: cwa-att-noneg-d1/test.parquet
  - config_name: cwa-att-noneg-d2
    data_files:
      - split: train
        path: cwa-att-noneg-d2/train.parquet
      - split: val
        path: cwa-att-noneg-d2/val.parquet
      - split: test
        path: cwa-att-noneg-d2/test.parquet
  - config_name: cwa-att-noneg-d3
    data_files:
      - split: train
        path: cwa-att-noneg-d3/train.parquet
      - split: val
        path: cwa-att-noneg-d3/val.parquet
      - split: test
        path: cwa-att-noneg-d3/test.parquet
  - config_name: cwa-att-noneg-d5
    data_files:
      - split: train
        path: cwa-att-noneg-d5/train.parquet
      - split: val
        path: cwa-att-noneg-d5/val.parquet
      - split: test
        path: cwa-att-noneg-d5/test.parquet
  - config_name: cwa-att-neg-d1
    data_files:
      - split: train
        path: cwa-att-neg-d1/train.parquet
      - split: val
        path: cwa-att-neg-d1/val.parquet
      - split: test
        path: cwa-att-neg-d1/test.parquet
  - config_name: cwa-att-neg-d2
    data_files:
      - split: train
        path: cwa-att-neg-d2/train.parquet
      - split: val
        path: cwa-att-neg-d2/val.parquet
      - split: test
        path: cwa-att-neg-d2/test.parquet
  - config_name: cwa-att-neg-d3
    data_files:
      - split: train
        path: cwa-att-neg-d3/train.parquet
      - split: val
        path: cwa-att-neg-d3/val.parquet
      - split: test
        path: cwa-att-neg-d3/test.parquet
  - config_name: cwa-att-neg-d5
    data_files:
      - split: train
        path: cwa-att-neg-d5/train.parquet
      - split: val
        path: cwa-att-neg-d5/val.parquet
      - split: test
        path: cwa-att-neg-d5/test.parquet
  - config_name: cwa-rel-noneg-d1
    data_files:
      - split: train
        path: cwa-rel-noneg-d1/train.parquet
      - split: val
        path: cwa-rel-noneg-d1/val.parquet
      - split: test
        path: cwa-rel-noneg-d1/test.parquet
  - config_name: cwa-rel-noneg-d2
    data_files:
      - split: train
        path: cwa-rel-noneg-d2/train.parquet
      - split: val
        path: cwa-rel-noneg-d2/val.parquet
      - split: test
        path: cwa-rel-noneg-d2/test.parquet
  - config_name: cwa-rel-noneg-d3
    data_files:
      - split: train
        path: cwa-rel-noneg-d3/train.parquet
      - split: val
        path: cwa-rel-noneg-d3/val.parquet
      - split: test
        path: cwa-rel-noneg-d3/test.parquet
  - config_name: cwa-rel-noneg-d5
    data_files:
      - split: train
        path: cwa-rel-noneg-d5/train.parquet
      - split: val
        path: cwa-rel-noneg-d5/val.parquet
      - split: test
        path: cwa-rel-noneg-d5/test.parquet
  - config_name: cwa-rel-neg-d1
    data_files:
      - split: train
        path: cwa-rel-neg-d1/train.parquet
      - split: val
        path: cwa-rel-neg-d1/val.parquet
      - split: test
        path: cwa-rel-neg-d1/test.parquet
  - config_name: cwa-rel-neg-d2
    data_files:
      - split: train
        path: cwa-rel-neg-d2/train.parquet
      - split: val
        path: cwa-rel-neg-d2/val.parquet
      - split: test
        path: cwa-rel-neg-d2/test.parquet
  - config_name: cwa-rel-neg-d3
    data_files:
      - split: train
        path: cwa-rel-neg-d3/train.parquet
      - split: val
        path: cwa-rel-neg-d3/val.parquet
      - split: test
        path: cwa-rel-neg-d3/test.parquet
  - config_name: cwa-rel-neg-d5
    data_files:
      - split: train
        path: cwa-rel-neg-d5/train.parquet
      - split: val
        path: cwa-rel-neg-d5/val.parquet
      - split: test
        path: cwa-rel-neg-d5/test.parquet

ProofWriter (The Cut)

A selection of proofwriter-mirror, not a transform of it: every row is a mirror row restricted to one question, with the same schema and every value verbatim. What this dataset adds is which rows and in what order, so a number measured on it can be read.

What was kept

Of ProofWriter's six question strategies, two — proof and rconc — the only two the whole task ladder can score, since a proof question has a derivation to recover and an rconc question's answer is that none exists.

The other four are dropped, and the reason is measurable. In an open-world, negation-free theory no negated atom is derivable, so every negative-polarity question is unentailed: audited across all 40 mirror configs, the entailed rate on those rows is 0.000 in all ten owa-noneg ones. Since the three inv-* strategies are exactly the negated half of the question set, half of any raw split is answerable by "if the statement says «not», answer False", with no reasoning at all. It cannot be balanced away — there is no entailed negative-polarity row to balance against — so it is trimmed.

random goes too: its failure chain is the bare (FAIL) at depth 0, a lookup refusal rather than a search. Dropping it drops depth 0 with it, and with that the corpus's heaviest skew.

How much, at which depth

Classes are matched per depth, so each split is 50 / 50 entailed / unentailed and chance is exactly 0.5. Depth is first-class on both sides: for an unprovable question ProofWriter's QDep is the depth the failed search reached — it equals the failure.depth witness on every unprovable row of all 40 configs — so an rconc row at depth k demands the same traversal as a proof row at depth k.

The corpus thins geometrically with depth, so an untapered split would be three-quarters shallow and its aggregate would nearly restate depth 1. Allocation walks depths deepest-first: the deepest cell is taken whole, and each shallower one gets 1.5x its deeper neighbour's realized count, floored and capped by availability.

The deep end is therefore bounded by the shelf, not by the recipe — on owa-att-noneg-d5, val holds 8 pairs at depth 5 against 40 at depth 1. Read a per-depth result with its n, and do not weigh the deepest point like the shallowest.

Splits, and the order

split	drawn from	meant for
train	mirror train	training
val	mirror dev	the curve during a run
test	mirror test	the reported number, read once

All three share one allocation: a validation curve measures the training distribution only if both are drawn from the same one.

The row order is part of the dataset. Rows are round-robined over (depth, class), so every cell contributes once before any contributes twice and the classes stay within one row of each other at every prefix — which means taking the first N rows is a fair sample at any N. Consumers that truncate without shuffling (a training run's num_examples, say) are why it is published this way rather than left to chance.

Which rows fill a cell depends on nothing but the corpus's own ids: the least-used theory first, ties broken by (theory_id, question_id). No seed, and no dependence on row order upstream.

• Derived from: proofwriter-mirror (itself a typed mirror of AI2's ProofWriter V2020.12.3).
• License: apache-2.0, inherited.
• Not affiliated with or endorsed by AI2; bundled here with attribution.

Citation

ProofWriter — and the RuleTaker corpora it builds on — are AI2's work. Please cite them:

@inproceedings{tafjord-etal-2021-proofwriter,
    title = "{P}roof{W}riter: Generating Implications, Proofs, and Abductive
             Statements over Natural Language",
    author = "Tafjord, Oyvind and Dalvi, Bhavana and Clark, Peter",
    booktitle = "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
    year = "2021",
    eprint = "2012.13048",
    archivePrefix = "arXiv",
}

@inproceedings{clark-etal-2020-transformers,
    title = "Transformers as Soft Reasoners over Language",
    author = "Clark, Peter and Tafjord, Oyvind and Richardson, Kyle",
    booktitle = "Proceedings of the 29th International Joint Conference on
                 Artificial Intelligence (IJCAI-20)",
    year = "2020",
    eprint = "2002.05867",
    archivePrefix = "arXiv",
}