Skip to content

Bug: get_c4 broken with datasets>=2.14 (allenai--c4 config removed) #87

Description

@UriKialy

Title: get_c4 broken with datasets>=2.14allenai--c4 config removed

Body:


Bug

lib/data.py get_c4() fails with recent versions of the datasets library (>=2.14):

ValueError: BuilderConfig 'allenai--c4' not found. Available: ['en', 'en.noblocklist', ...]

The config name allenai--c4 with explicit data_files no longer works. The fix is to use the standard 'en' config with streaming:

Before (broken)

traindata = load_dataset('allenai/c4', 'allenai--c4',
data_files={'train': 'en/c4-train.00000-of-01024.json.gz'}, split='train')

fix

traindata = list(load_dataset('allenai/c4', 'en', split='train', streaming=True).take(10000))

Same for validation split. Also valdata[:1100]['text'] needs to become [d['text'] for d in valdata] since streaming returns dicts not a Dataset object.

Additional: position_embeddings in newer transformers

lib/prune.py layer forward calls fail with transformers>=4.45 because LlamaDecoderLayer.forward() now requires position_embeddings as an explicit argument:

TypeError: cannot unpack non-iterable NoneType object

Fix: pre-compute RoPE embeddings and pass them to each layer call.

Environment

  • datasets==3.x
  • transformers==4.50+
  • Python 3.11

Happy to open a PR with the fixes if helpful.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

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