empirischtech/Palace-Config-Repair

Palace Configuration Repair Benchmark

Execution-graded repair tasks for configuration files of Palace, an open-source finite-element solver for computational electromagnetics. Each task gives a model a perturbed Palace JSON configuration and asks it to return a corrected one. A repair counts as correct only if it conforms to the schema, is accepted by the solver, and reproduces the reference outputs of the original case when Palace v0.14.0 runs it. A configuration that is valid and runs but simulates a different physical system fails.

Released with Beyond Schema Conformance with Executable Verification for Scientific Configuration Repair (Masood and Noor, NeurIPS 2026 Workshop on AI for Verifiable Coding, VeriCodeGen). The QLoRA adapter evaluated in the paper is empirischtech/Qwen2.5-Coder-32B-Palace-LoRA. It was fine-tuned on a separate corpus of 3,296 training conversations (plus 199 for validation) synthesized from Palace's documentation, schema options, and example configurations, not on these repair tasks. That corpus contains the three shipped reference configurations underlying 𝒟_ship and none of those underlying 𝒟_new.

Structure

Subset (config_name)	Split	Rows	Contents
reference (𝒟ref)	ship	3	Reference configurations shipped with Palace v0.14.0 and represented in the fine-tuning corpus (case-exposed)
reference (𝒟ref)	new	3	Newly authored reference configurations, created after the model and retrieval index were frozen (case-held-out)
perturbed (𝒟pert)	ship (𝒟ship)	54	3 cases × 6 perturbation classes × 3 variants
perturbed (𝒟pert)	new (𝒟new)	54	3 cases × 6 perturbation classes × 3 variants

Perturbed means altered relative to the benchmark target. A perturbed configuration is not necessarily schema-invalid or rejected by the solver.

from datasets import load_dataset

tasks = load_dataset("empirischtech/Palace-Config-Repair", "perturbed")   # splits: ship, new
refs = load_dataset("empirischtech/Palace-Config-Repair", "reference")   # splits: ship, new

Configurations are stored as JSON strings (json.loads them), because some perturbations deliberately change field types. Meshes and reference outputs are plain files under cases/:

cases/<case>/
  <config>.json      unperturbed reference configuration
  mesh/*.msh         mesh referenced by Model.Mesh
  ref/*.csv          Palace v0.14.0 reference outputs of that configuration (used for T2)
  gen_mesh.py        (new cases) gmsh script that generated the mesh
  verify.py          (new cases) closed-form capacitance check of ref/terminal-C.csv

Cases

Case	Split	Problem type	Model.L0	Reference CSVs
spheres	ship	Electrostatic	1e-2	8
rings	ship	Magnetostatic	1e-6	8
coaxial (coaxial_open.json)	ship	Transient	1e-3	4
parallel_plate	new	Electrostatic	1	7
layered_capacitor	new	Electrostatic	1	7
concentric_sphere	new	Electrostatic	1	7

The newly authored cases have closed-form capacitances. The flat cases match theirs to 5.4e-10 relative error and the faceted sphere to 0.51%. These are sanity checks only: grading always uses the Palace-generated CSVs in ref/.

Perturbation classes

Each perturbation is a deterministic edit of a reference configuration, applied by the same code to both splits. variant (0–2) indexes three fixed variants per (case, class); they are not random seeds.

error_class	Change made to the configuration	First failed outcome
drop_required_section	remove a required section	H0
wrong_type	replace a number with a string	H0
bad_enum	replace an allowed value with an invalid value	H0 or H1
unit_slip	multiply Model.L0 by 10³, 10⁻³, or 10⁶ (variants 0, 1, 2)	H2
dangling_ref	reference an identifier absent from the mesh	H2
multi_error	combine two perturbations	varies

All 6 reference configurations pass H0, H1, and H2, and all 108 perturbed configurations fail H2 before repair. Two perturbed configurations in 𝒟_new pass raw T2 while failing T0, which is why the outcomes are cumulative (see Grading).

Fields

reference

Field	Type	Description
case	string	case name; joins to perturbed.case
subset	string	ship or new
problem_type	string	Palace Problem.Type
description	string	one-line physical description
source	string	where the reference configuration comes from
in_training_corpus	bool	whether the configuration occurs in the paper's fine-tuning corpus (true for ship)
length_unit_L0	float	Model.L0, the mesh length unit in metres
config	string	reference configuration, JSON
config_path, mesh_path, reference_dir	string	paths into cases/
reference_files	list[string]	reference CSV filenames that grading compares

perturbed

Field	Type	Description
task_id	string	<case>__<error_class>__s<variant>
subset	string	ship or new
case, error_class, variant	string, string, int	task coordinates
prompt	string	exact model-facing repair request: a generic instruction plus the perturbed configuration. It does not reveal the perturbation class, the fix, or the target outputs.
perturbed_config	string	perturbed configuration, JSON (same content as in prompt)
perturbation_location	string	edited path(s) in the configuration
perturbation_original, perturbation_corrupted	string	JSON-encoded value before/after the edit
perturbation_description	string	human-readable ground truth for the edit. Do not show it to the model.
alters_L0	bool	true for the 12 tasks per split that modify Model.L0; these are specification-insufficient (see Caveats)
reference_config_path, reference_dir	string	reference configuration and reference outputs for grading

Grading

Grading uses Palace v0.14.0. Later releases changed the schema and the location of the regression references, so do not grade against them. Write the candidate configuration into a copy of cases/<case>/, under the reference configuration's filename, so that Model.Mesh resolves. Then record three raw indicators:

• T0: the configuration validates against Palace v0.14.0's JSON schema (scripts/schema/).
• T1: palace --dry-run <config> succeeds.
• T2: palace -np 1 <config> completes and its CSV outputs conform to ref/. Every reference CSV must be present, with matching header sets and row counts, unique header names, and finite numeric cells. Every value must satisfy |y − y*| ≤ 1e-12 + 1e-3·|y*|. error-indicators.csv, an adaptive-mesh-refinement heuristic, must be present and well-formed, but its values are not graded.

Report the cumulative outcomes H0 = T0 (structural conformance), H1 = T0 ∧ T1 (operational acceptance), and H2 = T0 ∧ T1 ∧ T2 (reference conformance). A repair is correct iff it reaches H2. H2 is an instance-level check of the selected outputs, not a formal proof of correctness.

Caveats

• Specification-insufficient tasks. alters_L0 marks 12 tasks per split that modify Model.L0: the 9 unit_slip tasks and the 3 multi_error variant-1 tasks, which combine a length-scale change with another perturbation. The prompt neither states the correct L0 nor gives enough information to determine it, so several physical scales are compatible with what the model sees (Suff(τ) = 0 in the paper). An H2 failure on these tasks does not by itself indicate weak repair. As in the paper, report H2 on all 54 tasks per split and on the 42 filtered tasks that do not modify Model.L0.
• Case exposure. 𝒟_ship is derived from Palace's public example configurations, which occur in the paper's fine-tuning corpus and may also appear in the pretraining data of general-purpose models. The corpus also contains repair exercises built from the same shipped configurations: 19 training and 6 validation conversations from spheres, rings, and coaxial_open, plus 19 training conversations from sibling coaxial configurations. They cover similar kinds of error, such as invalid enum values and missing sections. No benchmark task appears verbatim in the training or validation data. The reference configurations underlying 𝒟_new were authored after the model and retrieval index were frozen and have no exact configuration or case-identifier matches in the training, validation, or retrieval data. That check cannot rule out generic conceptual overlap. The paper's retrieval index contains only documentation and schema records, so no reference configuration from either split is retrievable at inference time.
• Few underlying cases. Each split derives its 54 tasks from only three cases, so results are evidence from this benchmark rather than population-level estimates. All three newly authored cases are electrostatic, while the shipped cases cover electrostatic, magnetostatic, and transient problems.

License

Apache-2.0. The shipped-case configurations, meshes, mesh scripts, and reference CSVs are redistributed from Palace v0.14.0, © Amazon.com, Inc. or its affiliates, under Apache-2.0; see cases/LICENSE-palace and cases/NOTICE-palace. The perturbed configurations in 𝒟_ship are modified versions of those files, with each change recorded in the perturbation_* fields.

Citation

@inproceedings{masood2026beyond,
  title     = {Beyond Schema Conformance with Executable Verification for Scientific Configuration Repair},
  author    = {Masood, Wasif and Noor, Shoaib},
  booktitle = {NeurIPS 2026 Workshop on AI for Verifiable Coding (VeriCodeGen)},
  year      = {2026}
}