com-kotobalabs/typed-decisions-code-holes

typed-decisions: code-holes

Single-token substitutions mined from the git history of 64 public repositories in the kotoba-lang organisation, each turned into one choice question with a gold answer: the token the commit actually put there. Part of kotoba-lang/typed-decisions (README 第8反復), whose trained model is com-kotobalabs/open-jev-deberta-v3-large. Sibling dataset (ungoaled): com-kotobalabs/typed-decisions-repo-governance.

What a record is

state = the commit message, the module's test namespace when one exists at that commit, and the module before the commit with exactly one token replaced by <<HOLE>>. questions[0] is a choice whose options are every same-kind token (keyword / symbol / whitespace-free string) already present in the module or its test, minus the token being replaced, and whose gold is the index of the token the commit put there. No option is invented; a hole whose gold was not among those tokens was dropped (68% of what was mined — the rename target is usually not in the file yet).

meta: repo, sha, src_path, old (the replaced token), kind, has_test, verified ("test" if the pair is one of the 24 FAIL→PASS verified fix pairs, else "commit"), n_options, and changed_tokens — inserted + deleted tokens in that file's whole diff.

Filter on changed_tokens. A single-token substitution inside a 500-token rewrite is a diff alignment, not a decision anyone made. Measured on the first held-out sample: 167/200 records had changed_tokens > 100, and a strong decision model scored 0.05 on their symbol holes; on changed_tokens ≤ 100 the same model scored 0.68 on keywords and 0.40 on symbols (chance 0.013 / 0.006).

Split

By repository (sha256(repo) % 10): test repositories never appear in train.

split	records	of which changed_tokens ≤ 100
train	877	273
validation	70	18
test	285	37

One repository from the mining run (app-kotoba-cloud, 68 records) is private and was removed before publishing; the counts above are after removal. Regenerate the full set from any checkout of the organisation with python -m typed_decisions.hole_data.

Measured with TypeSafe Jev (reference decision model, not trained on this)

344 isolated holes (changed_tokens ≤ 100, all splits, including the private repo's records): overall top-1 0.558; keyword 0.683 (n=183), string 0.50 (n=30), symbol 0.397 (n=131). A 0.8 confidence gate auto-applies 34% of holes at 0.92 precision. Hits on symbols are dependency rewires (clojure.string → kotoba.lang.text); misses are operator semantics (inc → quot) and namespace-prefix lookalikes. Full rows: reports/hole-eval-jev-iso100.json in the source repo.

Config repo-pool (added 2026-09-20)

Same mining, 18 public repositories × 150 commits, but the candidate pool is every source file in the repository at that commit (what a symbol index would offer), role-first capped at 255. Reach triples (22% → 64% of mined holes); meta.file_reachable says whether the module+test pool would have reached the gold too. 1,160 records. Measured with Jev on the 355 isolated holes: holes only the repo pool reaches score keyword 0.73 / string 0.69 / symbol 0.19; on the same holes the wider pool costs symbols about 3 in 26 versus the file pool while keywords are unchanged.

License

Apache-2.0. Source excerpts are from public repositories under their own licenses (kotoba-lang repositories are Apache-2.0 / MIT); this dataset carries the commit SHA and path for each excerpt.

Publisher

Published by Mithril (formerly Kotoba Cloud), operated by Kotoba Labs Inc. The Hugging Face handle com-kotobalabs is unchanged. Support: support.mithril.fund.