Use the GitHub Page or the repository links. If Colab does not open directly, open the notebook on GitHub first, then replace:
https://github.com/floriancafiero/lrec2026-llm-as-annotator-tutorial/blob/main/notebooks/NOTEBOOK.ipynb
with:
https://colab.research.google.com/github/floriancafiero/lrec2026-llm-as-annotator-tutorial/blob/main/notebooks/NOTEBOOK.ipynb
Click Connect in the top-right corner of Colab. If connection fails, reload the page and try again.
That is expected. The notebooks default to:
USE_API = False
In this mode, they use deterministic fallback predictions and do not call an external model.
Run notebooks in order:
00_setup_and_data.ipynb01_prompting_zero_few_shot.ipynb02_structured_outputs_and_validation.ipynb03_evaluation_and_error_analysis.ipynb04_sampling_and_bootstrapping.ipynbNotebook 00 creates the sample data used by later notebooks.
Re-run the previous cells from the top of the notebook. Colab runtimes are temporary and do not preserve all variables after restart.
This is part of the tutorial. Invalid JSON is one of the failure modes we want to detect. Do not manually fix it before validation unless the exercise explicitly asks you to.
This means the model added, removed, split, merged, translated, transliterated, or reordered tokens. Linguistic evaluation should not be computed on misaligned outputs.
Use File > Save a copy in Drive for your own edits, or reload the original notebook from the repository.
Replace the toy dataset with a CSV that has at least:
id, language, script, domain, text, tokens
For evaluation, also provide gold annotations such as:
gold_pos, gold_lemma, gold_morph
Switch back to:
USE_API = False
The rest of the workflow still works with fallback predictions.
In Colab, files are written to the temporary runtime unless you mount Google Drive. For the tutorial, temporary outputs are enough. For real work, save outputs to Drive or download them at the end of the session.