README.md

metadata

license: mit
task_categories:
  - question-answering
language:
  - en
pretty_name: General Knowledge Final Training Data

General Knowledge Final Training Data

This dataset bundle contains the training data used for the final general-knowledge checkpoint.

Contents

• stage1_targeted/train.jsonl: targeted supervised training data.
• stage1_targeted/validation.jsonl: held-out validation split for the targeted data.
• stage1_targeted/manifest.json: source and task counts for stage 1.
• stage2_direct_preference/train.jsonl: direct-preference training pairs built from actual model mistakes.
• stage2_direct_preference/validation.jsonl: held-out preference validation split.
• stage2_direct_preference/manifest.json: preference-pair construction summary.

The final preference stage uses only non-stress training/validation data. The public stress evaluation file is not included in this training bundle.