HaomingLuo/GPT-6-FEniCSx-Single-Attempt-Pass-Rate

GPT-6 FEniCSx Single-Attempt Pass Rate and Solver Corpus

A community evaluation and reusable scientific-code corpus by Haoming Luo.

A collection of 496 original, unedited first-attempt Python/FEniCSx solvers generated by GPT-6 Astra through the Codex CLI and passing the original PDEAgent-Bench accuracy and runtime gates in our evaluation environment.

Of 645 requested benchmark tasks, 644 have been evaluated: 496/644 = 77.02% first-attempt success. One task remains blocked by a reference-solver failure. This is an evaluation snapshot, not a completed 645-task leaderboard submission.

Contents

• successful_solutions.jsonl: successful first-attempt source code and measurements.
• solver_cases.zip: the same original Python files with inputs and predictions; source SHA-256 hashes are recorded in the data tables.
• results.jsonl: status records for all 645 tasks, including unsuccessful first attempts and the unevaluated reference-failure task. No repaired results enter this release's success rate.
• metadata.json: runtime and generation configuration.

Protocol

GPT-6 Astra (gpt-6-astra, high reasoning effort) generated one fresh response per task through Codex CLI 0.155.0-alpha.9.2. Tool use and internet access were disabled during generation; an external controller executed the returned code. Thus this measures first-response solver generation through Codex, not an unrestricted interactive Codex debugging session. Generation and solver budgets were each 300 seconds. Execution used one CPU and 4 GiB of memory, Python 3.12, DOLFINx 0.10.0.post2 and UFL 2025.2.

The original benchmark source terms, reference solutions and error/time gates were retained. The benchmark's two-sided NaN mask was also retained; finite_reference_coverage is provided separately. Benchmark success does not by itself establish complete spatial coverage or engineering validation. These are public benchmark tasks, not a held-out generalization claim.

Use

The code exposes solve(case_spec). Reproduction requires the corresponding benchmark task specification and scientific runtime. See REPRODUCE.md for execution instructions. Generated code and our helpers are licensed under Apache-2.0; evaluation records under CC BY 4.0 (see LICENSE.md). Successful solutions are a filtered subset; use the complete status table for evaluation-rate calculations. If these examples are used for training, their source benchmark must not subsequently be treated as unseen evaluation data.

Attribution

Based on PDEAgent-Bench, commit 0ba9853f82a78196796fa4eeaf0951eb4c000a00. Benchmark task IDs are preserved; upstream benchmark source files are not redistributed in this draft.

Downloadable numerical cases

solver_cases.zip contains the 496 unchanged Python modules, their runtime input specifications, shared helper, and saved predicted fields. Each case retains its original benchmark ID. checksums.json records archive/data hashes. These are benchmark-passing cases, not claims of validation for arbitrary engineering applications. We found no GPT-6 entry in the official published leaderboard data at the time this snapshot was prepared.