Datasets:
Datasheet
Composition
The resource contains five scientific reasoning benchmarks, four core protocol fixed evaluator where documented. Source problem records, generated messages, outcomes, and deterministic interventional labels are separated into tables.
Collection and filtering
LAB-Bench contains the project-defined llm_strict and text_no_tool text-only
slices. The broader slice must not be described as tool, retrieval, figure, or
supplementary-file use. Other benchmark slices follow the source files recorded
in the problem and run registries. Run coverage distinguishes the complete
source scope from the larger publication target, so filtered or pending cells
cannot be mistaken for a complete matrix.
ProtocolQA, and 600 SeqQA problems. FigQA and TableQA are excluded because they
carry required image assets. See docs/labbench_filter_audit.md.
Limitations
Generated traces can contain model errors and long-form reasoning that should not be treated as authoritative scientific advice. Runs outside the exact publication matrix remain provisional until frozen source manifests and exact problem-set validation are available. Request timestamps are unavailable, and token accounting for unmeasured cells remains null rather than reconstructed. The source staging includes legacy cost fields for supplemental GPT-OSS and Gemma-3 slices. Because those slices lack validated per-trajectory accounting and cost manifests, the release builder suppresses all six fields and records the exact suppression count in the release manifest. Client-observed wall time can be confounded by backend and run-batch conditions; in particular, recovery cells on a different backend must not be read as clean protocol-only latency comparisons.
Recommended use
Use the resource for evaluation analysis, protocol comparison, trace browsing, and research on interventional contribution labels. Honor each benchmark's license and contamination policy before any training use.