Datasets:

NDe2V
/

methods2test_small_cleaned

language:
  - en
license: mit
task_categories:
  - text-generation
tags:
  - code
  - java
  - unit-testing
  - methods2test
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*

methods2test_small_cleaned

A structurally-vacuous-filtered copy of the train split of andstor/methods2test_small (context config fm+fc+c+m+f+t+tc, the one actually used to fine-tune models in andstor/peft-unit-test-generation-replication-package). Produced for the investigation in lhnam/PEFT — FINDINGS.md (FINDINGS.md §1.2, §4 item 2), which found that 17.8% of the real fine-tuning targets are structurally vacuous (no assertion, empty, or tautological) and hypothesized this is a driver of the "convergence attractor" collapse seen when fine-tuning code LLMs for JUnit test generation.

What changed vs. the original

Split	Original rows	This dataset	Vacuous rate
`train`	7,440	6,124 (vacuous rows dropped)	17.7% removed
`validation`	953	953 (unchanged)	15.2% (left in, for fair eval_loss)
`test`	1,017	1,017 (unchanged)	16.9% (left in)

Only train is filtered. validation and test are byte-identical to the source dataset's fm+fc+c+m+f+t+tc config — the point of this dataset is to isolate the effect of training on cleaner targets while still measuring eval_loss / benchmark success against the real, unfiltered data distribution. Filtering only the split a model actually learns from, and leaving evaluation untouched, is what makes a before/after comparison causally meaningful.

Filtering method

Each target (the reference JUnit test) is classified as vacuous if it does not contain a real, non-tautological assert*/fail/verify call:

ASSERT_RE = re.compile(r"\b(assert\w*|fail|verify\w*)\s*\(", re.IGNORECASE)
TAUTOLOGY_RE = re.compile(
    r"assert(true)\s*\(\s*true\s*[,)]|assert(false)\s*\(\s*false\s*[,)]|"
    r'assertequals\s*\(\s*([A-Za-z0-9_."\']+)\s*,\s*\3\s*[,)]',
    re.IGNORECASE,
)

Targets under 15 characters are also treated as vacuous ("empty"). This is the exact classifier used throughout the source investigation (see scripts/filter_vacuous_training_data.py in the repo above), applied here with --mode drop.

Breakdown of the original train split before filtering:

Label	Count	%
`has_real_assert` (kept)	6,124	82.3%
`no_assert`	1,278	17.2%
`tautological_assert`	25	0.3%
`empty`	13	0.2%
vacuous total (dropped)	1,316	17.7%

(Matches FINDINGS.md's independently-reported 17.8% to within rounding — recomputed directly from this dataset's own source parquet.)

Columns

id (string) — original row id from andstor/methods2test_small.
source (string) — the prompt/context (unchanged).
target (string) — the reference JUnit test (the fine-tuning label).

No weight column — this is the drop variant, not downweight. See the source script if you want a down-weighted variant instead.

Intended use

Point a fine-tuning run's TRAIN_DATASET at this repo (config default) in place of andstor/methods2test_small (fm+fc+c+m+f+t+tc), keeping everything else — model, LoRA config, epochs, learning rate, validation split — identical, to test whether removing the training-time shortcut narrows or removes the post-fine-tuning "convergence attractor" documented in the source repo's FINDINGS.md. This is one experiment in an ongoing, self-correcting investigation — see that document for the full methodology, caveats, and history of revisions before citing any number from this dataset card in a paper.

Provenance

Source dataset: andstor/methods2test_small, config fm+fc+c+m+f+t+tc, revision confirmed via that dataset's own commit history.
Source paper / replication package: andstor/peft-unit-test-generation-replication-package.
License inherited as MIT from the source dataset.