Datasets:

Salesforce
/

BehaviorBench

# Croissant validation

Generated with:

```bash
python scripts/generate_croissant.py
```

Validated locally with the official `mlcroissant` CLI:

```bash
mlcroissant validate --jsonld=croissant.json --debug
```

Result: **PASS** with no errors or warnings. The metadata describes 30 Parquet distributions, 30 record sets and 625 validator-safe scalar/simple-array fields. Nested Parquet struct columns remain in the distributed files but are omitted from RecordSet field declarations because current mlcroissant record generation cannot reliably cast them.

Local artifact consistency was also checked with:

```bash
python scripts/validate_artifact.py
```

Result: **PASS**. All described local Parquet files exist, their schemas match the Croissant fields, required RAI keys are present, and local Markdown links resolve.

After uploading to `Salesforce/BehaviorBench`, rerun the NeurIPS online validator against the final public URLs and upload the validated JSON to OpenReview.

	# Croissant validation

	Generated with:

	```bash
	python scripts/generate_croissant.py
	```

	Validated locally with the official `mlcroissant` CLI:

	```bash
	mlcroissant validate --jsonld=croissant.json --debug
	```

	Result: PASS with no errors or warnings. The metadata describes 30 Parquet distributions, 30 record sets and 625 validator-safe scalar/simple-array fields. Nested Parquet struct columns remain in the distributed files but are omitted from RecordSet field declarations because current mlcroissant record generation cannot reliably cast them.

	Local artifact consistency was also checked with:

	```bash
	python scripts/validate_artifact.py
	```

	Result: PASS. All described local Parquet files exist, their schemas match the Croissant fields, required RAI keys are present, and local Markdown links resolve.

	After uploading to `Salesforce/BehaviorBench`, rerun the NeurIPS online validator against the final public URLs and upload the validated JSON to OpenReview.