Viewer jobs fail with JobManagerCrashedError for small-row-group Parquet
@lhoestq — could you inspect the failed viewer jobs for noetherlabs/probebench and requeue them if appropriate?
Dataset: https://huggingface.co/datasets/noetherlabs/probebench
Revision: efdf1b95a17a993f9adda1b2863da43c809ee4a6
Observed: 27 October 2026, 00:05 UTC
The dataset is public and ungated. The default/public_tuning split and first 100 rows return HTTP 200. The validity endpoint reports preview=true and viewer=false. An arbitrary row request returns HTTP 501 with JobManagerCrashedError:
https://datasets-server.huggingface.co/rows?dataset=noetherlabs%2Fprobebench&config=default&split=public_tuning&offset=100&length=1
The parquet and info endpoints report failed config-parquet/config-info jobs with no pending jobs. The card explicitly selects data/*.parquet. The dataset contains 32 Parquet shards; a local footer audit found 256 row groups with a maximum declared uncompressed row-group size of 812,706 bytes. There are 30,912 image rows. Separate raw-array archives are outside the configured Parquet paths.
Please let us know whether a packaging change is needed or whether the backend jobs can be retried. We can supply a minimal reproduction if the backend logs identify a specific failing shard or schema field.