Request write access to contribute evaluation artifacts for 22 runs (9 models) and a labeled scaled-time-limit experiment
Hi maintainers,
Could you grant Hugging Face user Yakundur the minimum write access needed to contribute evaluation artifacts to MatrAIx2026/PersonaCodeBench, or let us know your preferred pull-request submission workflow? A gated-dataset access request has also been submitted.
We have results for the 99 selected CPU tasks (upstream main 5933dd2, harness 0.5.0rc9, gpt-6-luna at medium effort as user simulator and judge) for 22 runs: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna and GPT-5.6 Terra at high / xhigh / max, Claude Fable 5.1 at high / xhigh / max, GPT-5.5 at high / xhigh, and GLM 5.3 and Grok 4.7 at high. We would like to upload them with one directory per model and effort, following the repository's existing artifact format.
We also ran a separate, non-standard experiment in which every interaction limit was scaled 2x and 3x (GPT-6 Astra and Claude Fable 5.1 at max effort). It would be clearly labeled as such and kept apart from the standard results.
The contribution will include scores, per-task results, coding-agent / simulator / judge call logs and relevant harness logs, with credentials and local machine paths removed; workspace copies and binaries excluded. We will preserve original error classifications, disclose deviations (simulator and judge served through a different API provider than the README's Azure deployment, calibration-gate retries, our rerun policy, and a separately labeled format-tolerant variant), and retain all previous attempts; 99 recorded tasks will not be described as 99 valid completions.
Thank you!