ThuraAung1601/reform-dafny-hint-gen

ReForm Dafny hint infill (DafnyBench-style)

A DafnyBench-style version of ThuraAung1601/reform-dafny-loop-inv-gen (itself derived from Veri-Code/ReForm-Python2Dafny-Dataset and Veri-Code/ReForm-DafnyComp-Benchmark).

The source task removes only loop invariants. Here, as in DafnyBench, every line starting with invariant, assert or decreases is removed from the verified program, so the model has to recover loop invariants, intermediate assertions and termination measures. The removal is the same line-based step DafnyBench used (it reproduces DafnyBench's own hints_removed exactly for 731 of 733 cleaned tasks), including its quirk of keeping the continuation lines of a multi-line assert (59 python2dafny rows), so training and benchmark tasks are built the same way.

file	config / split	rows	use
python2dafny_hint_infill.json	python2dafny / train	8,032	training
dafnycomp_hint_infill.json	dafnycomp / test	265	out-of-domain evaluation only

Removed hint lines in python2dafny: 39,148 invariant, 4,345 assert, 4,254 decreases.

Decontaminated against DafnyBench. 147 python2dafny programs overlapping a DafnyBench program (≥ 80% of the smaller program's 8-token shingles, identifiers ignored, so renamed copies count) were removed. The ids and their matching benchmark tasks are in build_report.json.

Fields

• id: source row id
• modified_input: the program with its invariant / assert / decreases lines removed (the task)
• output: the complete program, verified with dafny verify in the source dataset (unchanged)

License

CC-BY-NC-4.0, inherited from the source ReForm datasets -- non-commercial use, attribution required. Cite the original ReForm project.