Datasets:

anacodicai-labs
/

terrier-ta-benchmark

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
chunk_id: string
doc_id: string
chunk_index: int64
chunk_type: string
is_table: bool
is_figure: bool
text: string
embed_text: string
token_count: int64
section_path: string
pages: list<item: null>
  child 0, item: null
metadata: struct<book_title: string, subject: string, book_uuid: string, version: string, licence: string, pag (... 13 chars omitted)
  child 0, book_title: string
  child 1, subject: string
  child 2, book_uuid: string
  child 3, version: string
  child 4, licence: string
  child 5, page_id: string
course_id: string
source_file: string
module: string
section: string
page: null
task_a_course_qa: struct<books: int64, chunks: int64, courses: int64, token_count: struct<min: int64, max: int64, mean (... 745 chars omitted)
  child 0, books: int64
  child 1, chunks: int64
  child 2, courses: int64
  child 3, token_count: struct<min: int64, max: int64, mean: double, median: int64, over_512: int64>
      child 0, min: int64
      child 1, max: int64
      child 2, mean: double
      child 3, median: int64
      child 4, over_512: int64
  child 4, chunk_type: struct<prose: int64>
      child 0, prose: int64
  child 5, with_section_path: int64
  child 6, gold: struct<answerable: int64, abstention: int64, match_split: struct<fuzzy: int64, exact: int64>, negati (... 402 chars omitted)
      child 0, answerable: int64
      child 1, abstention: int64
      child 2, match_split: struct<fuzzy: int64, exact: int64>
          child 0, fuzzy: int64
          child 1, exact: int6
...
, cross_subject: int64
      child 4, per_course: struct<business-ethics: int64, foundations-of-information-systems: int64, introduction-to-business-2 (... 225 chars omitted)
          child 0, business-ethics: int64
          child 1, foundations-of-information-systems: int64
          child 2, introduction-to-business-2e: int64
          child 3, introduction-to-computer-science: int64
          child 4, introduction-to-python-programming: int64
          child 5, introductory-statistics: int64
          child 6, principles-of-data-science: int64
          child 7, principles-of-management: int64
          child 8, principles-of-marketing: int64
  child 7, licences: struct<Creative Commons Attribution-NonCommercial-ShareAlike License: int64, Creative Commons Attrib (... 21 chars omitted)
      child 0, Creative Commons Attribution-NonCommercial-ShareAlike License: int64
      child 1, Creative Commons Attribution License: int64
task_b_grading: struct<items: int64, questions: int64, score_scale: string, human_ceiling: struct<pearson_r: double, (... 96 chars omitted)
  child 0, items: int64
  child 1, questions: int64
  child 2, score_scale: string
  child 3, human_ceiling: struct<pearson_r: double, icc_2_1: double, mae_between_graders: double, exact_agreement: double, wit (... 20 chars omitted)
      child 0, pearson_r: double
      child 1, icc_2_1: double
      child 2, mae_between_graders: double
      child 3, exact_agreement: double
      child 4, within_1_point: double
to
{'task_a_course_qa': {'books': Value('int64'), 'chunks': Value('int64'), 'courses': Value('int64'), 'token_count': {'min': Value('int64'), 'max': Value('int64'), 'mean': Value('float64'), 'median': Value('int64'), 'over_512': Value('int64')}, 'chunk_type': {'prose': Value('int64')}, 'with_section_path': Value('int64'), 'gold': {'answerable': Value('int64'), 'abstention': Value('int64'), 'match_split': {'fuzzy': Value('int64'), 'exact': Value('int64')}, 'negative_types': {'sibling_subject': Value('int64'), 'cross_subject': Value('int64')}, 'per_course': {'business-ethics': Value('int64'), 'foundations-of-information-systems': Value('int64'), 'introduction-to-business-2e': Value('int64'), 'introduction-to-computer-science': Value('int64'), 'introduction-to-python-programming': Value('int64'), 'introductory-statistics': Value('int64'), 'principles-of-data-science': Value('int64'), 'principles-of-management': Value('int64'), 'principles-of-marketing': Value('int64')}}, 'licences': {'Creative Commons Attribution-NonCommercial-ShareAlike License': Value('int64'), 'Creative Commons Attribution License': Value('int64')}}, 'task_b_grading': {'items': Value('int64'), 'questions': Value('int64'), 'score_scale': Value('string'), 'human_ceiling': {'pearson_r': Value('float64'), 'icc_2_1': Value('float64'), 'mae_between_graders': Value('float64'), 'exact_agreement': Value('float64'), 'within_1_point': Value('float64')}}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              chunk_id: string
              doc_id: string
              chunk_index: int64
              chunk_type: string
              is_table: bool
              is_figure: bool
              text: string
              embed_text: string
              token_count: int64
              section_path: string
              pages: list<item: null>
                child 0, item: null
              metadata: struct<book_title: string, subject: string, book_uuid: string, version: string, licence: string, pag (... 13 chars omitted)
                child 0, book_title: string
                child 1, subject: string
                child 2, book_uuid: string
                child 3, version: string
                child 4, licence: string
                child 5, page_id: string
              course_id: string
              source_file: string
              module: string
              section: string
              page: null
              task_a_course_qa: struct<books: int64, chunks: int64, courses: int64, token_count: struct<min: int64, max: int64, mean (... 745 chars omitted)
                child 0, books: int64
                child 1, chunks: int64
                child 2, courses: int64
                child 3, token_count: struct<min: int64, max: int64, mean: double, median: int64, over_512: int64>
                    child 0, min: int64
                    child 1, max: int64
                    child 2, mean: double
                    child 3, median: int64
                    child 4, over_512: int64
                child 4, chunk_type: struct<prose: int64>
                    child 0, prose: int64
                child 5, with_section_path: int64
                child 6, gold: struct<answerable: int64, abstention: int64, match_split: struct<fuzzy: int64, exact: int64>, negati (... 402 chars omitted)
                    child 0, answerable: int64
                    child 1, abstention: int64
                    child 2, match_split: struct<fuzzy: int64, exact: int64>
                        child 0, fuzzy: int64
                        child 1, exact: int6
              ...
              , cross_subject: int64
                    child 4, per_course: struct<business-ethics: int64, foundations-of-information-systems: int64, introduction-to-business-2 (... 225 chars omitted)
                        child 0, business-ethics: int64
                        child 1, foundations-of-information-systems: int64
                        child 2, introduction-to-business-2e: int64
                        child 3, introduction-to-computer-science: int64
                        child 4, introduction-to-python-programming: int64
                        child 5, introductory-statistics: int64
                        child 6, principles-of-data-science: int64
                        child 7, principles-of-management: int64
                        child 8, principles-of-marketing: int64
                child 7, licences: struct<Creative Commons Attribution-NonCommercial-ShareAlike License: int64, Creative Commons Attrib (... 21 chars omitted)
                    child 0, Creative Commons Attribution-NonCommercial-ShareAlike License: int64
                    child 1, Creative Commons Attribution License: int64
              task_b_grading: struct<items: int64, questions: int64, score_scale: string, human_ceiling: struct<pearson_r: double, (... 96 chars omitted)
                child 0, items: int64
                child 1, questions: int64
                child 2, score_scale: string
                child 3, human_ceiling: struct<pearson_r: double, icc_2_1: double, mae_between_graders: double, exact_agreement: double, wit (... 20 chars omitted)
                    child 0, pearson_r: double
                    child 1, icc_2_1: double
                    child 2, mae_between_graders: double
                    child 3, exact_agreement: double
                    child 4, within_1_point: double
              to
              {'task_a_course_qa': {'books': Value('int64'), 'chunks': Value('int64'), 'courses': Value('int64'), 'token_count': {'min': Value('int64'), 'max': Value('int64'), 'mean': Value('float64'), 'median': Value('int64'), 'over_512': Value('int64')}, 'chunk_type': {'prose': Value('int64')}, 'with_section_path': Value('int64'), 'gold': {'answerable': Value('int64'), 'abstention': Value('int64'), 'match_split': {'fuzzy': Value('int64'), 'exact': Value('int64')}, 'negative_types': {'sibling_subject': Value('int64'), 'cross_subject': Value('int64')}, 'per_course': {'business-ethics': Value('int64'), 'foundations-of-information-systems': Value('int64'), 'introduction-to-business-2e': Value('int64'), 'introduction-to-computer-science': Value('int64'), 'introduction-to-python-programming': Value('int64'), 'introductory-statistics': Value('int64'), 'principles-of-data-science': Value('int64'), 'principles-of-management': Value('int64'), 'principles-of-marketing': Value('int64')}}, 'licences': {'Creative Commons Attribution-NonCommercial-ShareAlike License': Value('int64'), 'Creative Commons Attribution License': Value('int64')}}, 'task_b_grading': {'items': Value('int64'), 'questions': Value('int64'), 'score_scale': Value('string'), 'human_ceiling': {'pearson_r': Value('float64'), 'icc_2_1': Value('float64'), 'mae_between_graders': Value('float64'), 'exact_agreement': Value('float64'), 'within_1_point': Value('float64')}}}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

TerrierTA Benchmark

Two tasks for course-grounded educational AI:

Task A — course-grounded retrieval and abstention. A question about a course must be answered from that course's own material, with a citation. A question whose answer is not in the course must be abstained on.
Task B — rubric-grounded short-answer grading. Grade a student's free-text answer against an instructor reference answer, and compare against two independent human graders.

Task A

Corpus: 9 OpenStax university textbooks (9 courses), parsed from OpenStax's structured content rather than PDFs, so there is no parse loss and every chunk carries a chapter/section breadcrumb.

Book	Subject	Pages	Chunks	Licence
Introduction to Computer Science	computer-science	204	886	Creative Commons Attribution-NonCommercial-ShareAlike License
Principles of Data Science	computer-science	112	572	Creative Commons Attribution-NonCommercial-ShareAlike License
Introduction to Python Programming	computer-science	131	263	Creative Commons Attribution-NonCommercial-ShareAlike License
Foundations of Information Systems	computer-science	132	540	Creative Commons Attribution-NonCommercial-ShareAlike License
Introduction to Business 2e	business	267	704	Creative Commons Attribution-NonCommercial-ShareAlike License
Principles of Management	business	240	648	Creative Commons Attribution-NonCommercial-ShareAlike License
Principles of Marketing	business	311	632	Creative Commons Attribution-NonCommercial-ShareAlike License
Business Ethics	business	113	421	Creative Commons Attribution-NonCommercial-ShareAlike License
Introductory Statistics	statistics	200	571	Creative Commons Attribution License

Chunks: 5237 (token counts: min 5, median 467, mean 430.5, max 624; 603 above 512). Sized with tiktoken for text-embedding-3-small at 512 tokens.
With a section breadcrumb: 5237 of 5237.
Gold questions: 527 answerable (span match: {'fuzzy': 58, 'exact': 469}) and 501 abstention items ({'sibling_subject': 238, 'cross_subject': 263}).

Each answerable question was drafted from one specific chunk and its answer verified as a span of that chunk, so gold_chunk_ids is correct by construction; items whose answer could not be matched were dropped.

Abstention items re-ask an answerable question against a different course. They come in two difficulty tiers, because random cross-course negatives are trivially easy:

sibling_subject — a different course in the same subject (heavy vocabulary overlap).
cross_subject — a course in a different subject (the easy baseline).

Files

task_a_course_qa/manifest.csv              one row per book, with its licence
task_a_course_qa/chunks.jsonl              one chunk per line (schema below)
task_a_course_qa/gold/questions.jsonl      id, question, gold_answer, course_id,
                                           gold_chunk_ids, section_path, match,
                                           expected (answer|abstain), negative_type
task_a_course_qa/raw_pages/<course>/*.json master page records — re-chunk from these
                                           without refetching

Chunk schema: chunk_id, doc_id, chunk_index, chunk_type, is_table, is_figure, text, embed_text, token_count, section_path, pages, metadata. text is the citation/display text; embed_text is what an embedder receives (section breadcrumb prepended).

Task B

Source: the UNT short-answer grading corpus (Mohler, Bunescu and Mihalcea, ACL-HLT 2011) — an introductory computer-science course, with every student answer graded independently by two people.

Items: 2273 graded student answers across 81 questions, scored 0-5.
Human ceiling (the two graders against each other):

Metric	Value
Pearson r	0.5974
ICC(2,1)	0.4989
MAE between graders	0.7481
Exact agreement	0.5557
Within 1 point	0.7937

Two trained graders agree exactly on roughly half of the answers. A system should therefore be reported against this ceiling, not against perfect agreement.

The student answers are not redistributed here

The source corpus ships with no licence, only a citation request, so this repository does not host its student answers. Rebuild them locally:

python task_b_grading/fetch_mohler.py     # downloads the official archive, writes items.jsonl

This repository ships only the aggregate human_ceiling.json and the build scripts.

Scoring rules that matter

gold_chunk_ids fit these chunks. If you re-chunk at a different size, match on the gold answer span instead of chunk ids.
Report abstention alongside retrieval. A system that answers everything scores well on recall and fails abstention; report both, and report the two negative tiers separately.
State the embedding model, its dimension, and the generator model next to every number.

Licensing and attribution

This dataset is released under CC BY-NC-SA 4.0, inherited from the OpenStax source texts (most of the selected books are CC BY-NC-SA; ShareAlike propagates to derivatives). It is therefore non-commercial.

OpenStax textbooks © Rice University, reused under their Creative Commons terms; per-book licences are recorded in task_a_course_qa/manifest.csv.
Task B derives from the UNT corpus; if you use it, cite:

@inproceedings{mohler2011learning,
  author    = {Michael Mohler and Razvan Bunescu and Rada Mihalcea},
  title     = {Learning to Grade Short Answer Questions using Semantic Similarity
               Measures and Dependency Graph Alignments},
  booktitle = {ACL-HLT 2011},
  year      = {2011}
}

Limitations

Task A questions are model-drafted and then span-verified against their source chunk; a human-validated subsample is planned rather than complete.
Task B is a single course with a small cohort. The grades are real and double-marked, but the domain is narrow.
The corpus is English-language university material in computer science, business and statistics.

Downloads last month: 81

Total file size:

54.5 MB