FailDiag: Selected Program Failure Diagnosis Benchmark

Companion dataset for FailDiag: Evaluating LLMs for Diagnosis of Selected Program Failures by Huy Tran and Thanh Thi Nguyen.

Dataset summary

FailDiag contains 600 real C++ selected-failure diagnosis cases drawn from 110 programming tasks:

• 300 Programming Fundamentals (PF) cases
• 300 Data Structures and Algorithms (DSA) cases
• 110 tasks (55 PF, 55 DSA)

Each case provides the faulty program as submitted, one selected failure as an expected-versus-observed pair (or a runtime error), same-lineage repair evidence (the author's own later passing revision, the diff between them, and the post-repair test outcomes), and a structured diagnostic reference.

What the benchmark measures

A program may contain multiple genuine defects. FailDiag evaluates whether a diagnosis identifies the defect that explains the selected failure, rather than merely identifying any defect in the program. A response that correctly describes a real but unrelated bug is not counted as correct.

Task statements

field	meaning
task_statement_original	the exact task text used in the reported paper experiments, byte for byte
task_statement_en	an English-only public-facing representation, for accessibility and future evaluations

Exact reproduction of the reported experiments MUST use task_statement_original. task_statement_en is a different input representation unless byte-identical to the original, which english_is_byte_identical_to_original records per task (55 of 110).

PF statements are bilingual Vietnamese/English in the source; DSA statements are English. The English rendering starts as an exact slice of the original at its own English separator; any further change is recorded per task.

translation_source	tasks
official_english	98
official_english_with_translated_missing_parts	6
official_english_with_source_correction	6
translated_from_vietnamese	0

A small number of English renderings contain documented corrections of obvious inconsistencies in the original bilingual source, while the original evaluated text remains unchanged. The 6 affected tasks are PF06, PF39, PF40, PF42, PF43, PF46.

Four tasks have no statement in the source. DSA07, DSA37, DSA40, DSA50 carry the placeholder Problem statement unavailable in the retained source artifact., marked task_statement_status = "placeholder_in_source". Nothing was invented for them.

Evidence conditions

condition	evidence
Code-Only (CO)	task statement + faulty program
Failure-Informed (FI)	CO + selected failing fixture/input, expected behaviour, observed behaviour or runtime error
Repair-Assisted (RA)	FI + canonical same-lineage repair evidence

The public case records intentionally contain the fields needed for all three conditions. Construct condition-specific inputs; do not feed a complete record to every condition.

Loading evidence safely

The archival public JSONL contains the evidence needed to construct all three experimental conditions. Passing an entire case record directly to a diagnosis model would leak evidence across conditions — the selected failure into Code-Only, the repair into Failure-Informed, and the structured reference into all three. Use the bundled reference loader.

from faildiag import load_cases, load_tasks, get_case, build_case_evidence

cases = load_cases()
tasks = load_tasks()
case = cases[0]                       # or get_case("<case_id>")

co = build_case_evidence(case, "CO", tasks=tasks)
fi = build_case_evidence(case, "FI", tasks=tasks)
ra = build_case_evidence(case, "RA", tasks=tasks)

From the command line:

python faildiag.py --case-id <CASE_ID> --condition CO
python faildiag.py --case-id <CASE_ID> --condition FI
python faildiag.py --case-id <CASE_ID> --condition RA
python faildiag.py --case-id <CASE_ID> --condition CO --task-statement-field task_statement_en

Evidence hierarchy:

• CO = task + faulty program
• FI = CO + selected-failure evidence
• RA = FI + same-lineage repair evidence

Structured references are evaluation targets and are never part of diagnosis-model evidence. The loader excludes root concept, defect-form status, atomic defect form, causal location, reference diagnosis and causal chain from every packet, in every condition.

For exact reproduction of the paper, use task_statement_original — the loader's default. task_statement_en is available through the explicit task_statement_field argument and is not exact reproduction of the reported experiments.

The loader ships with tests and an exhaustive audit over all 600 cases: 0 / 0 / 0 cross-condition leakage violations for CO / FI / RA (reports/loader_leakage_audit.json).

Structured outputs

A diagnosis is scored on four structured fields:

1. Root concept — one of 13 values.
2. Defect-form status — one of ATOMIC, COMPOSITE, UNSCOREABLE.
3. Atomic defect form — one of 9 values, applicable only when the status is ATOMIC.
4. Failure-relevant code location — visible-line intervals over the displayed faulty program (N, N-M, or a comma-separated list). The loader's line_numbered=True option renders the program with the L<n> | prefixes these intervals index.

Full value lists are in schema/faildiag_public_release_schema.json under structured_reference_vocabulary.

The structured reference set was created through an AI-assisted curation process with final human approval. A subsequent stratified blinded human-led audit of 120 cases was used to assess and refine reference reliability. The released structured references are the finalized benchmark references. Aggregate results from the 120-case blinded audit are provided in reports/reference_audit_summary.md. Per-case annotation-workflow provenance is internal and is not part of this release.

Dataset organization

benchmark/     faildiag_cases_public.jsonl   one record per case (600 records)
tasks/         faildiag_tasks.{jsonl,json,csv}  one record per task (110 records)
cases/         case_to_task.{csv,jsonl}      every case mapped to its task
faildiag.py    condition-safe reference loader and CLI
tests/         loader tests and the all-case leakage audit
schema/        field-by-field description of every record type
reports/       reference audit summary, translation and source-correction audit, privacy
               review, validation report, loader leakage audit
RELEASE_METADATA.json   authors, license, counts, vocabulary, reproducibility checksums
CITATION.cff   citation metadata
LICENSE        CC-BY-4.0
CHANGELOG.md   what the public-release transformation did
SHA256SUMS     digest of every file in this release

Task text is normalised at the task level and not duplicated across cases. Each case references its task_id and carries task_statement_original_sha256.

Example record

Abbreviated and synthetic — field shapes only, not a real submission:

{
  "case_id": "unified::<64-hex>",
  "display_order": 1,
  "domain": "PF",
  "task_id": "PF::<64-hex>",
  "task_key": "PF01",
  "task_title": "[Example] Sum of an array",
  "task_statement_original_sha256": "<64-hex>",
  "reference": {
    "root_concept": "INDEX_OR_RANGE",
    "defect_form_status": "ATOMIC",
    "defect_form": "WRONG_EXPRESSION",
    "causal_location": "3",
    "diagnosis": "The loop bound compares i <= n, so the loop reads one element past the end of the array.",
    "causal_chain": "loop bound uses <= n -> a[n] is read past the end of the array -> the returned sum includes an out-of-bounds value"
  },
  "evidence": {
    "language": "cpp",
    "failing_code": "int total(int* a, int n) {\n    int s = 0;\n    for (int i = 0; i <= n; i++) {\n        s += a[i];\n    }\n    return s;\n}",
    "repair_code": "int total(int* a, int n) {\n    int s = 0;\n    for (int i = 0; i < n; i++) {\n        s += a[i];\n    }\n    return s;\n}",
    "repair_diff": "@@ -3,1 +3,1 @@\n-    for (int i = 0; i <= n; i++) {\n+    for (int i = 0; i < n; i++) {",
    "repair_validation_summary": {
      "repair_shape": "SINGLE_SUBSTANTIVE_EDIT",
      "successor_full_pass": true
    },
    "selected_failing_test": {
      "test_id": "t3",
      "test_index": 3,
      "input": "3\n1 2 3",
      "fixture": "3\n1 2 3",
      "expected": "6",
      "actual_before": "32774",
      "error_before": null,
      "actual_after": "6",
      "passed_after": true
    },
    "visible_failing_code_line_count": 7
  }
}

causal_location is a string of visible-line intervals over failing_code; here "3" is the loop-bound line, and visible_failing_code_line_count matches the seven visible lines. causal_chain is a string, and defect_form is one of the nine atomic forms.

## Reproducing the paper

1. use **`task_statement_original`**;
2. construct **CO / FI / RA** with the bundled loader, or according to the documented
   evidence conditions;
3. use the paper's prompt, output-schema and configuration artifacts;
4. **do not substitute `task_statement_en`** for the original task statement.

`RELEASE_METADATA.json` records `paper_evaluation_task_field = "task_statement_original"` and
the digest of the internal authoritative annotation file the reported experiments consumed.

## Known source issues

- **PF19** — the bilingual source declares the parameter as `int** matrix`
  in one half and `int**& matrix` in the other. Left **intentionally unresolved**.
- **DSA07, DSA37, DSA40, DSA50** — four DSA tasks carry a source placeholder instead of a statement.
- **PF06, PF39, PF40, PF42, PF43, PF46** — obvious source-English inconsistencies corrected **only in
  `task_statement_en`**; the original evaluated text is preserved unchanged.

Detail: `reports/translation_and_source_correction_audit.csv`.

## Privacy and provenance

Automated privacy screening found no direct identifiers under the implemented checks. Manual review was completed for flagged release items, but automated screening cannot guarantee the absence of all incidental identifiers.

The release excludes LMS account identifiers, student identity information, private student
linkage, submission-level restricted provenance, account histories and private salts, each
listed with its reason in `RELEASE_METADATA.json`.

**No claim of absolute anonymity is made.** The programs are real student submissions and may
contain incidental content automated screening cannot detect. Numeric literals carrying task
semantics, such as a modulus or a `#define` string literal, are deliberately not redacted.

## License

Released under the **Creative Commons Attribution 4.0 International** license
(**CC-BY-4.0**) — see [`LICENSE`](LICENSE) and <https://creativecommons.org/licenses/by/4.0/>.

You may share and adapt the material, including commercially, provided you give appropriate
credit, link to the license, and indicate if changes were made.

## Archival release and citation

The frozen public release associated with the reported FailDiag experiments is archived on
Zenodo at <https://doi.org/10.5281/zenodo.22269090>. The archive corresponds to Hugging Face revision
`0aa5c7a19daad4595fbc3dc43ffa03cf79a52861`.

- The **version DOI** [`10.5281/zenodo.22269090`](https://doi.org/10.5281/zenodo.22269090) identifies the exact frozen
  release: one archive whose contents are fixed and can never change.
- This **Hugging Face repository may receive later documentation updates** — corrections to
  the dataset card, citation metadata and reports. Its revision hash therefore moves over
  time, while the Zenodo archive does not.
- **Exact reproduction of the reported experiments should use the archived Zenodo version**,
  together with `task_statement_original` as described above.

A concept DOI, [`10.5281/zenodo.22269089`](https://doi.org/10.5281/zenodo.22269089), also exists and always resolves to the
most recent version. Use it to refer to the dataset in general; **do not cite it for exact
reproducibility**, because it does not identify a fixed snapshot.

```bibtex
@dataset{tran_faildiag_2026,
  author    = {Huy Tran and Thanh Thi Nguyen},
  title     = {FailDiag: Selected Program Failure Diagnosis Benchmark},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22269090},
  url       = {https://doi.org/10.5281/zenodo.22269090}
}

The associated manuscript, FailDiag: Evaluating LLMs for Diagnosis of Selected Program Failures, has no DOI assigned yet.

Machine-readable metadata is in CITATION.cff. Corresponding author: Thanh Thi Nguyen tnguyen5@usc.edu.au.