Skip to content

Parquet upload fails #104

Open
Open
@platypii

Description

@platypii

I generated a parquet file by using the tokenize_data.py script in the together-python repo to generate a parquet file with tokenized data. I followed the steps here https://docs.together.ai/docs/fine-tuning-data-preparation#tokenized-data

python tokenize_data.py --tokenizer openai-community/gpt2 --out-filename dataset.parquet

I then tried to upload this dataset to together:

import { upload } from 'together-ai/lib/upload'
await upload('dataset.parquet')
"failed to read parquet file dataset.parquet"

With a bit of added logging in together-typescript I found the more specific error message:

invalid parquet version

Which is coming from inside the parquetjs library which is woefully unmaintained in 5+ years and does not support the majority of modern parquet files. #102 also ran into a parquet parsing issue. Might be worth using a different parquet library.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions