Datasets:

turboblitz
/

textinguisher

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
id: string
doc: string
arxiv_id: string
year: int64
hist_year: int64
signature: string
first_error: string
missing_file: string
n_files: int64
engine: string
error_lines: list<item: struct<line: int64, text: string, file: string, fileLabel: string>>
  child 0, item: struct<line: int64, text: string, file: string, fileLabel: string>
      child 0, line: int64
      child 1, text: string
      child 2, file: string
      child 3, fileLabel: string
license: string
answer_score: int64
fix_source: string
fix_files: list<item: string>
  child 0, item: string
answer_url: string
lane: string
url: string
file_edits: list<item: struct<file: string, search: string, replace: string, type: string>>
  child 0, item: struct<file: string, search: string, replace: string, type: string>
      child 0, file: string
      child 1, search: string
      child 2, replace: string
      child 3, type: string
answer_id: int64
error_category: string
question_id: int64
prompt_v1: string
repo_url: string
mutation: string
title: string
question_author: struct<user_id: int64, name: string, url: string>
  child 0, user_id: int64
  child 1, name: string
  child 2, url: string
single_shot_visible: bool
parent_sha: string
error_message: string
prompt: string
repo: string
created: timestamp[s]
changed_files: list<item: string>
  child 0, item: string
companions: list<item: string>
  child 0, item: string
tags: list<item: string>
  child 0, item: string
question_url: string
answer_license: string
reported_file: string
bib: bool
fix_answer: string
log_tail: string
accepted_answer_id: int64
answer_author: struct<user_id: int64, name: string, url: string>
  child 0, user_id: int64
  child 1, name: string
  child 2, url: string
score: int64
source: string
original_root: string
sha: string
target: string
to
{'bib': Value('bool'), 'lane': Value('string'), 'file_edits': List({'file': Value('string'), 'search': Value('string'), 'replace': Value('string'), 'type': Value('string')}), 'reported_file': Value('string'), 'fix_files': List(Value('string')), 'single_shot_visible': Value('bool'), 'error_lines': List({'line': Value('int64'), 'text': Value('string'), 'file': Value('string'), 'fileLabel': Value('string')}), 'log_tail': Value('string'), 'prompt': Value('string'), 'target': Value('string'), 'id': Value('string'), 'doc': Value('string'), 'mutation': Value('string'), 'source': Value('string'), 'repo': Value('string'), 'sha': Value('string'), 'parent_sha': Value('string'), 'license': Value('string'), 'url': Value('string'), 'changed_files': List(Value('string')), 'original_root': Value('string'), 'n_files': Value('int64'), 'engine': Value('string'), 'error_category': Value('string'), 'prompt_v1': Value('string'), 'repo_url': Value('string'), 'fix_answer': Value('string'), 'answer_id': Value('int64'), 'answer_score': Value('int64'), 'fix_source': Value('string'), 'companions': List(Value('string')), 'question_id': Value('int64'), 'accepted_answer_id': Value('int64'), 'title': Value('string'), 'tags': List(Value('string')), 'score': Value('int64'), 'created': Value('timestamp[s]'), 'error_message': Value('string'), 'question_url': Value('string'), 'answer_url': Value('string'), 'question_author': {'user_id': Value('int64'), 'name': Value('string'), 'url': Value('string')}, 'answer_license': Value('string'), 'answer_author': {'user_id': Value('int64'), 'name': Value('string'), 'url': Value('string')}}
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 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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
              id: string
              doc: string
              arxiv_id: string
              year: int64
              hist_year: int64
              signature: string
              first_error: string
              missing_file: string
              n_files: int64
              engine: string
              error_lines: list<item: struct<line: int64, text: string, file: string, fileLabel: string>>
                child 0, item: struct<line: int64, text: string, file: string, fileLabel: string>
                    child 0, line: int64
                    child 1, text: string
                    child 2, file: string
                    child 3, fileLabel: string
              license: string
              answer_score: int64
              fix_source: string
              fix_files: list<item: string>
                child 0, item: string
              answer_url: string
              lane: string
              url: string
              file_edits: list<item: struct<file: string, search: string, replace: string, type: string>>
                child 0, item: struct<file: string, search: string, replace: string, type: string>
                    child 0, file: string
                    child 1, search: string
                    child 2, replace: string
                    child 3, type: string
              answer_id: int64
              error_category: string
              question_id: int64
              prompt_v1: string
              repo_url: string
              mutation: string
              title: string
              question_author: struct<user_id: int64, name: string, url: string>
                child 0, user_id: int64
                child 1, name: string
                child 2, url: string
              single_shot_visible: bool
              parent_sha: string
              error_message: string
              prompt: string
              repo: string
              created: timestamp[s]
              changed_files: list<item: string>
                child 0, item: string
              companions: list<item: string>
                child 0, item: string
              tags: list<item: string>
                child 0, item: string
              question_url: string
              answer_license: string
              reported_file: string
              bib: bool
              fix_answer: string
              log_tail: string
              accepted_answer_id: int64
              answer_author: struct<user_id: int64, name: string, url: string>
                child 0, user_id: int64
                child 1, name: string
                child 2, url: string
              score: int64
              source: string
              original_root: string
              sha: string
              target: string
              to
              {'bib': Value('bool'), 'lane': Value('string'), 'file_edits': List({'file': Value('string'), 'search': Value('string'), 'replace': Value('string'), 'type': Value('string')}), 'reported_file': Value('string'), 'fix_files': List(Value('string')), 'single_shot_visible': Value('bool'), 'error_lines': List({'line': Value('int64'), 'text': Value('string'), 'file': Value('string'), 'fileLabel': Value('string')}), 'log_tail': Value('string'), 'prompt': Value('string'), 'target': Value('string'), 'id': Value('string'), 'doc': Value('string'), 'mutation': Value('string'), 'source': Value('string'), 'repo': Value('string'), 'sha': Value('string'), 'parent_sha': Value('string'), 'license': Value('string'), 'url': Value('string'), 'changed_files': List(Value('string')), 'original_root': Value('string'), 'n_files': Value('int64'), 'engine': Value('string'), 'error_category': Value('string'), 'prompt_v1': Value('string'), 'repo_url': Value('string'), 'fix_answer': Value('string'), 'answer_id': Value('int64'), 'answer_score': Value('int64'), 'fix_source': Value('string'), 'companions': List(Value('string')), 'question_id': Value('int64'), 'accepted_answer_id': Value('int64'), 'title': Value('string'), 'tags': List(Value('string')), 'score': Value('int64'), 'created': Value('timestamp[s]'), 'error_message': Value('string'), 'question_url': Value('string'), 'answer_url': Value('string'), 'question_author': {'user_id': Value('int64'), 'name': Value('string'), 'url': Value('string')}, 'answer_license': Value('string'), 'answer_author': {'user_id': Value('int64'), 'name': Value('string'), 'url': Value('string')}}
              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.

TeXtinguisher: a benchmark of real LaTeX compile errors

Real LaTeX projects that fail to compile with a located error, paired (where a human fix exists) with the smallest SEARCH/REPLACE edit that makes them compile. Verifier: latexmk (TeX Live 2025/2026). Built by scripts/build_release.py; integrity in MANIFEST.json (rows, bytes, sha256 per file).

Files

file	rows	what
`texse_heldout.jsonl`	287	TeX.SE single-file benchmark (question's MWE + accepted answer's document)
`texse_train.jsonl`	922	training split of the same mining run
`texse2_heldout.jsonl`	669	TeX.SE relaxed mining (error mentioned in title/body; fix from accepted or top answer)
`texse2_train.jsonl`	2122	its training split
`texse_pool.jsonl` / `texse2_pool.jsonl`	1774 / 3923	unlabeled failing MWEs (no human fix; compile-reward only)
`multi2_heldout.jsonl` / `multi2_train.jsonl`	30 / 87	multi-file projects: TeX.SE questions with companion files + GitHub commits
`skew_eval.jsonl`, `skew_multi_heldout.jsonl`, `skew_train.jsonl`	184 / 200 / 1628	arXiv "toolchain skew" rows, metadata only (see below)
`refs.json`	7424 docs	PDF text + page count of the fixed document, keyed by `doc` (`pseudo: true` = teacher-generated, pool rows only)
`projects/<id>/`	9814 dirs	sources needed to reproduce the compile (one per unique id)

Row ids are unique; id names the project directory, doc the reference entry.

Row format

Common fields: id, doc, engine (pdflatex/xelatex/lualatex), error_lines (parsed {file, line, text}), license; non-skew rows add log_tail, prompt (the production fix prompt: error lines + source window + log tail), error_category. Skew rows are metadata only (no project directory, no prompt/log_tail/error_category): arxiv_id, year, hist_year, signature, first_error, missing_file, n_files.

Single-file rows: projects/<id>/main.tex is the fixed document; apply break_edits (replace search by replace, once, in order) to obtain the failing document. target is the fix as <edit> SEARCH/REPLACE blocks. Pool rows have no break_edits/target; their main.tex is the failing document.
Multi-file rows (lane: project): projects/<id>/ is the fixed project, file_edits break it, target uses <edit file="..."> blocks. prompt is the project-context prompt (file tree + reported file); prompt_v1 is the single-file prompt for the reported file only. reported_file, fix_files, changed_files, n_files describe the row.
TeX.SE provenance: question_id, question_url, title, tags, score, created, question_author ({user_id, name, url}), license; answer_id/accepted_answer_id, answer_url, answer_author, answer_license only when an answer was used (absent on 3641 pool rows).
GitHub provenance: repo, repo_url, sha (fixing commit), parent_sha (failing state), url (commit), license.

Local absolute paths were rewritten: the project path to projects/<id>, other /Users/<name>/ prefixes to ~/. Absolute paths that remain are content of the posts themselves (e.g. \graphicspath{{/home/...}}).

How rows were mined and verified

TeX.SE (scripts/mine_texse.py, mine_texse2.py, mine_texse_pool.py, mine_texse_multi.py): from the archive.org data dump stackexchange_20251231. A question qualifies if its body has a full \documentclass ... \end{document} MWE and reports an error (texse: a pasted ! ... line; texse2: an error mention in title/body). The MWE must fail locally with an error located in main.tex; the fix document (longest complete document in the accepted answer, else the top answer with score >= 2 for texse2; otherwise a snippet) must compile clean. Diff capped at 6 hunks / 40 lines; edits must apply uniquely and round-trip exactly. Rows whose first error is a missing file/font/graphic or an engine mismatch (fontspec, Unicode engine) were dropped, so those classes are absent. Funnels and error histograms: data/texse_report.md, texse2_report.md, texse_pool_report.md, multi2_report.md.
GitHub multi-file (scripts/mine_github_multi.py, mine_github_multi2.py): commit pairs where the parent fails with a located error and the child compiles, edits round-trip exactly, max 5 rows per repo. 13 github_multi rows were never re-verified locally and have no reference entry; they are not in this release (unlicensed anyway).
Splits: by question id (stratified by error category) for TeX.SE, by repo/question for multi-file, by arXiv id for skew. Held-out ids, question ids, repos and arXiv ids do not appear in any train or pool file (checked by build_release.py --check, which also reports near-duplicates: Jaccard >= 0.9 on normalized main.tex lines finds one, texse2_437213 vs pool row texse2_pool_437217, a follow-up question). Note that texse_heldout over-represents rows that were probed before the split (not a random sample).

Metrics (`scripts/run_eval.py`)

compiles (fixed): the model's edits apply and latexmk exits clean. Headline metric; the only one for pool and skew rows.
deletion-bounded (del<=40 in scripts/heldout_table.py): compiles and the edits remove at most 40 characters of non-comment content (fix_env.deleted_chars). A heuristic guard against "delete until it compiles"; it does not see content hidden by \iffalse or comment environments.
strict: compiles and the PDF text matches the reference (refs.json): same page count and word-level SequenceMatcher ratio >= 0.985. Penalizes valid alternative fixes.
exact: the patched source equals the human-fixed source. TeX.SE labels are whole answer documents, often restyled (28% of texse rows have >= 20 diff lines), so exact/strict under-count correct minimal fixes; filter on diff_lines for a minimal-fix subset.

Licences and attribution

source	licence	rows	attribution required
TeX.SE (`texse*`, multi `source: texse`)	CC BY-SA 2.5 / 3.0 / 4.0 per post (`ContentLicense` in the dump)	49 / 5014 / 4717	author name + profile link (`question_author`, `answer_author`) and post links (`question_url`, `answer_url`)
GitHub (`multi2*`, `source: github`)	MIT 17, GPL-3.0 4, Apache-2.0 4, CC0-1.0 4, AGPL-3.0 2, CC-BY-SA-4.0 2, CC-BY-4.0 1	34	per the repo's LICENSE; `repo_url`, `sha`, `license` per row
arXiv (`skew*`)	mostly arXiv non-exclusive-distrib 1.0 (1872 / 2012), some CC BY / BY-SA / BY-NC	2012	metadata only shipped

TeX.SE license is the question's; on 125 rows answer_license differs (78 newer, 47 older; each post carries its own ContentLicense). Fixed documents are derived from both posts: the TeX.SE portion is share-alike.

Excluded and why

github_forks rows (22; 7 were in multi2_heldout): students' theses / CVs on templates, no licence on the fork author's text. Also removes the 3 CJK rows that are unfixable under the eval harness's engine choice.
GitHub rows with license None or NOASSERTION (97 of 131 non-fork GitHub rows).
arXiv sources (skew*, multi_synth, project_*, eval, hard_eval*): arXiv's non-exclusive licence does not permit redistribution; skew rows keep the metadata fields listed under Row format.
project_train_all.jsonl, project_eval.jsonl (Sundial user projects), distilled/teacher files, and everything under corpus/ not referenced above.
refs.json for skew rows does not exist (no clean baseline compile by construction).

Regenerating skew rows

scripts/gen_skew.py --target 200 --seed 3 (arXiv 2008-2018 single-root papers compiled under TeX Live 2025) and scripts/gen_skew_train.py (2010-2020 multi-file papers, historical TeX Live) download the sources from arXiv and rebuild corpus/arxiv_skew*/<arxiv_id>/files; match rows by arxiv_id. Every skew row is an organic failure (no planted error), so there is no ground-truth patch and only compiles applies.

Contact

Sundial (see repository)

Downloads last month: -

Total file size:

216 MB