ThuraAung1601/dafnybench-multitask

DafnyBench multi-task

ThuraAung1601/dafnybench-cleaned turned into six Dafny tasks, for evaluation. All comments were removed from every program (the ReForm training programs have none); each ground truth was re-verified with Dafny 4.11.0 afterwards. 731 of 733 programs are used (skipped: 2 ground truth uses {:verify false}, 1 loop-invariants: ground truth fails its own faithfulness check, 1 termination: ground truth fails its own faithfulness check).

task	the model gets	answer is scored by
loop-invariants	the program with its loop invariants removed	dafny verify + only proof hints may differ
proof-annotations	every line starting with invariant / assert / decreases removed (DafnyBench's construction)	dafny verify + only those lines may differ
single-hint	exactly one invariant / assert / decreases clause removed, one the proof needs (checked)	dafny verify + only proof hints may differ
termination	every decreases clause removed, where Dafny can't infer them back (checked)	dafny verify + only proof hints may differ; no decreases *
proof-repair	one invariant / assert / decreases clause broken (deleted, a comparison's strictness flipped, a conjunct dropped or a literal changed) plus Dafny's error output (error column)	dafny verify + only proof hints may differ
spec-generation	every method's requires / ensures removed	dafny verify + Re:Form's subset reward: each generated precondition must be implied by the ground truth's, each generated postcondition must imply it (checked by Dafny: a check method assumes the ground-truth precondition, calls the method -- so only its generated spec and its frame are known -- and asserts the ground-truth postcondition)

task	rows
loop-invariants	469
proof-annotations	632
single-hint	475
termination	68
proof-repair	562
spec-generation	617

proof-annotations is DafnyBench's own task, rebuilt from the comment-free program with DafnyBench's line-based construction (so continuation lines of multi-line hints stay in the task, as in DafnyBench).

Fields

• id: <program id>__<task code> (li, pa, sh, tm, pr, sg)
• orig_id: the source program -- split by this, so a held-out program is in no task's training data
• task: the task name; its system prompt is PROMPTS[task] in dafny_tasks.py
• modified_input: the program the model gets
• output: the complete, verified program (the answer)
• error: Dafny's error output for proof-repair rows (part of the prompt: dafny_tasks.user_message), else empty
• meta: JSON: which clause was removed or broken, and how
• test_file: DafnyBench's file name

dafny_tasks.py holds the prompts, faithfulness checks and the subset check used to build and score this data.