noanabeshima's picture
Update README.md
8ada67f verified
metadata
license: mit
configs:
  - config_name: default
    data_files:
      - split: ratio1_v1
        path: ratio1_v1.csv
      - split: ratio1_v2
        path: ratio1_v2.csv
      - split: ratio2_v1
        path: ratio2_v2.csv
      - split: ratio3_v1
        path: ratio3_v2.csv
      - split: ratio6_v2
        path: ratio6_v2.csv
      - split: ratio10_v1
        path: ratio10_v1.csv
      - split: ratio30_v1
        path: ratio30_v1.csv
      - split: ratio50_v1
        path: ratio50_v1.csv
      - split: test
        path: test.csv

This dataset is composed of Claude-labelled fineweb documents.

For each document, Claude is asked if it is 'forecastable' (i.e. would be a reasonable seed for a pastcasting question) and to estimate the date the document was published. V1 splits were generated by having Claude label ~50K random fineweb documents and v2 splits were augmented with labels on ~30K additional documents that a DebertaV3 classifier finetuned on ratio10_v1 thought were forecastable (Claude thought ~1/3 of these additional documents were forecastable).

Splits are of the form ratio{negative_to_positive_ratio}_v{1 or 2}. For example ratio6_v2 is ~6 negative examples for each positive example. Splits do overlap.

ratio6_v2 was used to train https://huggingface.co/noanabeshima/forecastability-classifier-v1.

Prompt can be found in prompt.txt. It was iterated on slightly using a small set of ground-truth human labels. GPT-4.1 performed slightly better, but Claude had a more favorable TOS for open-sourcing data/models.

Claude considers ~2% of fineweb documents to be forecastable.

Made with the help of Collin Gray.