Creating JSONL format for Vertex AI batch inference from GCS files

Other Questions ai employee
1

I’m struggling with generating a proper JSONL file that includes references to images stored in my Google Cloud Storage bucket. I need this for running batch inference using Vertex AI.

What I’m trying to do:

  • Extract file paths from my GCS bucket
  • Format them appropriately for Vertex AI batch processing.

My current approach:

  1. First, I list all files: gsutil ls gs://my-data-bucket/images > file_list.txt
  2. Then, I manually convert the txt file to JSONL format like this:
{"content": "gs://my-data-bucket/images/photo1.jpg", "mimeType": "image/jpeg"}
{"content": "gs://my-data-bucket/images/photo2.jpg", "mimeType": "image/jpeg"}

The problem:
When I submit my batch prediction job, I keep getting an error indicating that the file “cannot be parsed as JSONL.”

I suspect there might be formatting issues with my JSONL structure. Has anyone faced this before? Is there a more straightforward way to export bucket contents into the correct JSONL format that Vertex AI requires?

2

Had this exact problem with my first batch inference pipeline. Turns out invisible characters from gsutil output were breaking the JSON parsing. I ditched gsutil and wrote a Python script using the GCS client library to generate the JSONL directly. The script loops through bucket objects and writes each line with proper JSON encoding - no more formatting issues. Also make sure your JSONL structure matches what your model wants - some need extra fields like “instances” wrapping the content. Once I switched to generating it programmatically, all the parsing errors went away.

3

yeah, also make sure each line in your jsonl is properly formatted, no extra spaces or weird chars. that could be causing errors too. try using a tool like jq to check it out!