JSON source data file too large
#1
by
albertvillanova
HF staff
- opened
This discussion is a follow-up of this issue: https://github.com/huggingface/datasets-server/issues/1215
As
@AntonioRenatoMontefusco
reported, currently the viewer shows a JobManagerCrashedError:
Job manager crashed while running this job (missing heartbeats).
The reason is that the datasets library tries to load into RAM memory the JSON source data file, and this has more than 4 GB.
One of the solutions is to change the format of the data files, from JSON to JSON-lines, so that the data files can be loaded in memory line by line, thus with manageable RAM requirements.
@AntonioRenatoMontefusco
uploaded a Json-lines file: train.jl
However there are two issues yet:
- First, the old JSON file (
train.json) should be deleted from the repo; otherwise thedatasetslibrary tries to load both, the JSON and the JSON-Lines files - Second, the extension of the JSON-Lines file should be
.jsonlinstead of.jl