基準評測

Terminal-Bench 2.1 (2.1) agentic terminal benchmark bundles. Task containers and verifier assets live in eval storage; Postgres stores only the manifest and pass/fail results.

GSM8K deterministic disjoint eval shards from openai/gsm8k/main:test. Running all 13 shards covers the full 1,319-question dataset; Postgres stores only shard metadata and pass/fail results.

HellaSwag deterministic disjoint eval shards from Rowan/hellaswag:validation. Running all 13 shards covers the full 10,042-question dataset; Postgres stores only shard metadata and pass/fail results.

ARC Challenge randomized eval shards from ai2_arc/ARC-Challenge:test. Question text lives in S3; Postgres stores only shard metadata and results.

HumanEval+ randomized eval shards from evalplus/humanevalplus:test. Question text lives in S3; Postgres stores only shard metadata and results.

MBPP+ randomized eval shards from evalplus/mbppplus:test. Question text lives in S3; Postgres stores only shard metadata and results.