---
license: apache-2.0
tags:
- code-similarity
- parallel-corpus
- competitive-programming
- code-retrieval
- C
- Rust
pretty_name: C-to-Rust Parallel Semantic Similarity Corpus
size_categories:
- 1K<n<10K
---

# Dataset Card for C-to-Rust Parallel Semantic Similarity Corpus

## Dataset Summary
The **C-to-Rust Parallel Semantic Similarity Corpus** is a curated dataset consisting of **1,886 aligned, function-level C and Rust code pairs**. It was developed to evaluate cross-language semantic similarity and functional equivalence between a traditional legacy language (C) and a modern memory-safe language (Rust). 

The source code snippets are drawn from accepted competitive programming submissions across **Project CodeNet**, **xCodeEval**, and **The Algorithms**. Every pair shares identical algorithmic logic, has been validated to compile under modern environments (GCC C17 and Rust 1.94.0), and is filtered to ensure functional correctness.

## Dataset Structure

The dataset is distributed as a single JSONL file where each row represents an aligned problem pair. 

### Data Fields
* `problem_id` (`string`): A zero-padded sequential identifier spanning `0001` to `1886`.
* `problem_description` (`string`): The full natural language problem statement provided by the source platform.
* `c_code` (`string`): The normalized, fully functional C implementation.
* `rust_code` (`string`): The normalized, fully functional Rust implementation.
* `difficulty` (`string`): The alignment difficulty tier assigned to the pair (`easy`, `medium`, or `hard`).

---

## Dataset Creation & Curation

### Filtering and Preprocessing Pipeline
1. **Functional Correctness:** Restricted exclusively to submissions marked as "Accepted" by online judges to guarantee true semantic alignment.
2. **Compilation Validation:** Submissions were strictly compiled locally using GCC (C17 standard) and Rust 1.94.0 to eliminate syntax errors or compiler drift.
3. **Code Normalization:** Dead code, redundant macros, comments, and explicit Rust `#[allow(...)]` attributes were systematically stripped. Code layouts were standardized using `clang-format` and `rustfmt`.
4. **Function Inlining:** Core logic was restricted to a single function block, and all function identifiers were normalized to `solution` to eliminate superficial retrieval shortcuts.
5. **Semantic Diversity:** The final corpus contains a diverse set of unique problems with no overlapping duplicates, providing a clean benchmark for cross-language evaluation.

### Difficulty Categorization
Difficulty tiers were empirically mapped using zero-shot cosine similarity scores from the `SFR-Embedding-Code-400M_R` model. Quantile thresholds at the 25th and 75th percentiles partition the dataset:
* **Easy (25%):** Similarity $\ge 0.868$
* **Medium (50%):** Similarity between $0.781$ and $0.868$
* **Hard (25%):** Similarity $< 0.781$ (Representing heavy syntax/paradigm divergence)

---
## Associated Paper

The complete academic paper detailing the methodology, curation pipeline, and evaluation results for this dataset is available directly within this repository:

* 📄 **Read the Paper:** [`c-rust-parallel-corpus.pdf`](./c-rust-parallel-corpus.pdf)

Please refer to the paper for in-depth insights into the dataset's design choices and baseline benchmarks.
---

## Citation Information

```bibtex
@proceedings{hejlek2026creating,
  title={Creating a Parallel C to Rust Corpus for Semantic Similarity Evaluation},
  author={Hejlek, Vojtěch},
  year={2026},
  organization={Instituto de Ciências Matemáticas e de Computação, Universidade de São Paulo (ICMC/USP)},
  note={Advisor: Alneu de Andrade Lopes, Coadvisor: Leonardo Jesus Almeida}
}