qa-formatting-data / README.md
nzhao721's picture
Add the OpenStax QA-formatting data
6d1b0e7 verified
|
Raw History Blame Contribute Delete
2.57 kB
metadata
license: cc-by-nc-sa-4.0
language:
  - en
pretty_name: 'Same Content, Different Format: QA-formatting data'
configs:
  - config_name: passages
    data_files: data/passages.jsonl
  - config_name: training_documents
    data_files: data/training_documents.jsonl
  - config_name: heldout_questions
    data_files: data/heldout_questions.jsonl
  - config_name: training_probe_questions
    data_files: data/training_probe_questions.jsonl
  - config_name: rewrites
    data_files: data/rewrites.jsonl
tags:
  - question-answering
  - knowledge-retrieval
  - openstax

Same Content, Different Format: QA-formatting data

Data for the QA-formatting experiment of Same Content, Different Format: QA Augmentation and Faded Worked Examples Under a Fixed Training Budget (OpenStax main study). The trained checkpoints are in nzhao721/qa-formatting-models.

Configs

Config Rows Contents
passages 1,962 the OpenStax passages, each with its source-grounded claims (5,886 in all) and two accepted answers per claim
training_documents 70,632 the training documents, 23,544 per arm; condition is q0_normal_fact, q1_diverse_prose or q3_diverse_qa
heldout_questions 11,774 new questions about trained claims (two to three per claim), never used in training
training_probe_questions 23,544 the QA arm's training questions, verbatim
rewrites 5,886 the accepted GPT-5.4 mini rewrites (four prose and four QA variants per claim)

reports/ holds the build, validation and generation reports and the 107 rejected rewrite records.

The normal-fact arm repeats each claim's supporting source span in four identical documents; the prose and QA arms use four rewritten variants per claim. The synthetic-data preflight's data is not included.

Licensing

All material derives from OpenStax textbooks (OpenStax, Rice University, https://openstax.org), obtained through HuggingFaceTB/openstax_paragraphs, with claims extracted from the passages and rewritten by a language model. OpenStax now licenses its textbooks under CC BY-NC-SA 4.0 and states that they may not be used to train large language models without its prior written permission, so this dataset is distributed under CC BY-NC-SA 4.0 (attribution, non-commercial use, adaptations under the same license). The per-row license field in passages records the license stated when the passages were collected and is superseded by this notice.