Skip to content

feat: add wave 2 code benchmark adapters - #3

Merged
aabbdev merged 6 commits into
mainfrom
research/bcb-arm64-20260806
Aug 7, 2026
Merged

feat: add wave 2 code benchmark adapters#3
aabbdev merged 6 commits into
mainfrom
research/bcb-arm64-20260806

Conversation

@aabbdev

@aabbdev aabbdev commented Aug 7, 2026

Copy link
Copy Markdown
Owner

Summary

  • add the pinned public LiveBench Coding 2024-11-25 task and benchmark-specific extraction/scoring semantics
  • add BigCodeBench v0.1.4 complete/instruct and full/hard surfaces with the official amd64 evaluator plus a reproducible source-equivalent arm64 Hard runtime
  • keep generated programs behind network-disabled, read-only, capability-dropped Docker sandboxes and fail closed when canonical BigCodeBench calibration cannot run
  • add source-backed protocol audits for OJBench, TIR-Bench, Codeforces ELO, Spider, VITA-Bench, and MEGA instead of exposing approximate scores

BigCodeBench arm64 evidence

The canonical Hard preflight improved from 79/148 to 145/148 after resolving the evaluator dependency set and routing writable caches to tmpfs. The remaining records (BigCodeBench/101, /590, and /1012) require live external downloads; they now produce canonical evaluation errors and no model score while networking remains disabled.

The arm64 runtime is therefore restricted to subset="hard". subset="full" requires the official runtime until a separate full-suite validation is completed.

Validation

  • pre-commit run --all-files
  • pytest — 464 passed, 11 skipped
  • targeted BigCodeBench mypy — success across 6 source files
  • canonical BigCodeBench Hard arm64 preflight — 145/148, 3 network-dependent fail-closed records
  • locked arm64 image rebuild — 143-package lock matches the built image
  • arm64 Docker Compose smoke — passed with network disabled/read-only filesystem
  • uv build — wheel and sdist built; Dockerfile, both compose files, and arm64 lock verified in artifacts

Known limitations

  • the pinned official linux/amd64 BigCodeBench image still needs a privileged native x86_64 Docker smoke; QEMU on Apple Silicon segfaulted and the available GPU instance disallowed layer registration/unshare
  • strict parity for the three network-dependent Hard records requires immutable audited fixtures, not enabling network for generated code

@github-actions

github-actions Bot commented Aug 7, 2026

Copy link
Copy Markdown

✅ Benchmark documentation has been automatically updated.

@aabbdev
aabbdev merged commit 8031b9a into main Aug 7, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant