CLEVDEV/icml_obf

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- code
- python
- leetcode
- codeforces
- iterative
- recursive
- obfuscation
size_categories:
- 1K<n<10K
---

# Iterative vs Recursive Code Pairs — with Full AST Obfuscation

Verified iterative/recursive Python solution pairs from LeetCode and Codeforces.
Extends `CLEVDEV/full_before_conv_icmll_pub` with full-variable AST
obfuscation applied to every user-defined identifier.

## Columns added on top of the source dataset

- `iterative_solution_fullobf` — iterative solution with every user-defined
variable, argument, function, class, and import renamed to `v{n}` / `f{n}`
tokens. Type annotations and docstrings stripped. Comments removed
(byproduct of `ast.unparse`).
- `recursive_solution_fullobf` — same treatment for the recursive variant.
Names shared across both variants map to the same token (single transformer
per problem).
- `fullobf_token_map` — JSON string `{original_name: obfuscated_token}` (the
unified map across both variants).
- `fullobf_status` — `verified` | `verification_failed` | `parse_error` |
`missing`. `verified` means BOTH variants pass all the original test cases
after obfuscation.
- `fullobf_iter_passed`, `fullobf_rec_passed` — per-variant pass flags.
- `fullobf_iter_num_passed`, `fullobf_rec_num_passed` — per-variant counts.

## What is protected (not renamed)

- Python builtins (`print`, `range`, `len`, etc.) and exceptions
- stdlib flat names (`Counter`, `defaultdict`, `heappush`, `inf`, …)
- `self`, `cls`, dunder names
- LeetCode harness requirements: `Solution`, `ListNode`, `TreeNode`, `Node`
- LeetCode entry-method argument names (the harness calls by kwarg name)
- Class-level attribute names (accessed via `self.X`)

## Source

Built from `CLEVDEV/full_before_conv_icmll_pub` using a custom
`ast.NodeTransformer`. Every row is verified against the same execution
harnesses used to produce the source dataset (LeetCode subprocess harness
for `class Solution`, Codeforces stdin/stdout harness for CF).