Datasets:
Tasks, models and repair attempts
A SimRepairBench task is a fixed damaged OpenUSD package with a repair brief and a versioned objective verifier. The same task can be given to different models, or repeated with the same model, without changing the input. Each attempt retains its trajectory, submitted candidate and grading evidence.
| Field | Identifies |
|---|---|
case_id |
The object lineage, such as R01 |
task_revision |
The immutable input, requirements and verifier version |
model_config_id |
A canonical hash of the provider, model, reasoning, generation, tool, prompt and harness configuration |
attempt_id |
The positive replicate number for that task and model configuration |
episode_id |
One execution, including an unsuccessful or interrupted execution |
campaign_id |
The accounting group that pays for the execution |
An attempt is unique within a campaign by task_revision, model_config_id and attempt_id. Case identifiers stay stable across versions. Changing the damaged package or verifier creates a new task revision; changing the model configuration creates a new model configuration identifier.
Follow an episode
The episode linkage schema connects the immutable input to its artifacts. Every artifact link records a confined path and SHA-256 digest. The input also records its package digest, USD digest, Hugging Face revision and Harbor task digest.
Each model receives a rendered image of the exact broken input alongside the repair brief. The image digest is part of the task identity and its bytes remain in the episode input. Candidate-only probes expose joint, collision and mass measurements that a still image cannot show. The model configuration records this observation policy.
The trajectory contains the model observations, actions and tool events. The candidate manifest identifies the files captured from the actor's submission directory, and its capture event links those bytes back to the trajectory. A failed execution may have no submitted candidate; that absence remains explicit.
The intact source is a grading reference. It is never inserted into an episode as a placeholder repair result. The native 3D gallery opens broken inputs; intact references and model submissions have separate USD links. Captured candidate archives retain their original bytes, while any browsable result layout keeps comparison thumbnails outside gallery discovery paths. Source reconstructions, equivalent-authoring examples and adversarial submissions used to qualify a verifier are controlled fixtures. They have no model episode identity and do not enter the repair results index.
Read the grades
Each episode links separate objective, visual and physical reports. The objective report describes the measured authored-USD criteria, per-defect recovery and preservation of healthy content. The visual report records the independent judge configuration and trusted rendered evidence. A physical report records simulation measurements when they are performed; not_run has a null score.
The analysis links these reports to the exact submitted candidate and explains any remaining faults. Scores do not replace the underlying evidence, and one grading method does not silently stand in for another. These links make the same fixed task comparable across models while keeping every submitted repair and its outcome retrievable.