--- license: mit pretty_name: ChipBench Verified tags: [verilog, rtl, hardware-design, benchmark] --- # ChipBench Verified A fixed version of [ChipBench](https://github.com/zhongkaiyu/ChipBench). Same format and file names as the original, so the original harness runs it as is. We checked each task's spec against its hidden tests: | Tasks | Spec and tests agree | Weak tests | Unfair tests | Both | |---|---|---|---|---| | 134 | 68 | 14 | 43 | 9 | Weak tests: the spec states behaviour the tests never check. Unfair tests: the tests require something the spec does not state. Both: the task has both problems. We fixed the spec in 52 tasks and broken tests in 36 tasks, and added checks for unchecked spec behaviour in 18 tasks. In 5 tasks with weak tests we did not add checks, so their tests can still miss a design that skips some stated behaviour. Every changed task was re-run to confirm it is passable from the spec. Diff against the original files to see each change. Layout: `Verilog Gen/dataset_*/` and `Verilog Debugging/dataset_debug_zero_shot_*/` with `ProbNNN__prompt.txt`, `_ref.sv` and `_test.sv`. Covers Verilog Gen (45) and zero-shot Verilog Debugging (89). More details: https://chipagents.ai/research/open-rtl-benchmarks Derived from ChipBench (https://github.com/zhongkaiyu/ChipBench) under its original license; see LICENSE.