Datasets:
Preserve MSE validation evidence for synthetic replay claims
#3
by booth-algo - opened
Curates the paired H100/Llama-3.1-8B/vLLM validation runs and keeps debug runs separate from the paper-facing table.
Constraint: Submission data needs a clean HF viewer config plus raw JSON provenance.
Rejected: Uploading every R2 MSE artifact | mixed smoke tests and superseded broad sweeps obscure the final claim.
Confidence: high
Scope-risk: narrow
Directive: Keep future MSE additions tiered into main vs supplementary and preserve pair metadata.
Tested: Parsed README YAML; generated 28-row summary.parquet and 96-row fidelity_deltas.csv; verified min success_rate >= 0.75 and SWE cascade deltas.
Not-tested: HF dataset loading via datasets library locally; package is not installed in this environment.
booth-algo changed pull request status to merged
booth-algo deleted the
refs/pr/3 ref