dhyuti-n/xlcost-variable-roles

XLCoST Variable Roles

Program-level code with structurally derived variable-role labels for probing how code LLMs represent variables. Built from XLCoST (Zhu et al., 2022).

Five roles, labeled from AST/structural analysis — never from the variable's name — so probes trained on these labels must rely on context:

Role	Definition
index_key	used as an array index or dict key (arr[i], d[key])
accumulator	target of +=-style updates or .append()-style calls inside a loop
iterator	bound in a loop header (for x in …, for (int i = …)
boolean	assigned a boolean literal
class_struct	declared class/struct name

Configs

python_perturbations — every Python program under 10 naming strategies: baseline, random_nouns, single_chars (a, b, c…), all_same (everything → x), numeric_vars (v1, v2…), and misleading_<role> for each role (role variables get counter-role names, all other variables get role-looking names). Role labels are re-extracted from the transformed code.

multilingual_baseline — original programs in all 7 XLCoST languages (C++, Java, Python, C#, Javascript, PHP, C) with role labels, for cross-language transfer experiments.

Fields

{
  "id": "10005:Python:baseline",
  "problem_id": 10005,
  "language": "Python",
  "split": "train",
  "strategy": "baseline",
  "code": "def maxPresum(a, b): ...",
  "roles": {
    "index_key": ["i"],
    "accumulator": ["X"],
    "iterator": ["i"],
    "boolean": [],
    "class_struct": []
  }
}

Rows store code plus role-name sets rather than token-level labels, so the dataset is model-agnostic: map names to token labels with whatever tokenizer you are probing (reference implementation in the companion pipeline's probing.label_tokens).

Labeling method

Python roles come from the ast module; the other six languages use regex extractors (subscripts, augmented assignments/increments/collector calls, loop headers, boolean assignments, class/struct declarations) with per-language keyword exclusion. PHP identifiers are labeled without the $ sigil. Programs with no role-labeled variable are dropped.

Provenance and credits

Source programs are the program-level nl2code_search release of XLCoST, which pairs GeeksforGeeks solutions across 7 languages. All credit for the underlying corpus goes to the XLCoST authors; this dataset adds only the role labels and the renaming variants. XLCoST is released under the Apache 2.0 license, as is this derivative.

@article{zhu2022xlcost,
  title   = {XLCoST: A Benchmark Dataset for Cross-lingual Code Intelligence},
  author  = {Zhu, Ming and Jain, Aneesh and Suresh, Karthik and
             Ravindran, Roshan and Tipirneni, Sindhu and Reddy, Chandan K.},
  journal = {arXiv preprint arXiv:2206.08474},
  year    = {2022}
}

If you use the role labels or perturbations, please also cite this dataset.

Known limitations

• Non-Python role labels are regex-derived and inherit that noise (e.g. ++ in a for-header counts toward accumulator, matching the original experimental protocol).
• XLCoST code is competitive-programming style; identifier names are already short and partially uninformative, which attenuates renaming effects.
• Renaming is whole-word textual substitution with keyword/builtin protection, not scope-aware alpha-renaming.
• Under all_same, distinct variables collapse to one name, so exclusion-based roles (notably accumulator, which excludes loop and index variables) survive in far fewer programs; this mirrors the original experimental protocol, where labels are re-extracted after renaming.