README.md

metadata

language:
  - en
  - zh
  - es
  - ur
license: apache-2.0
task_categories:
  - text-generation
tags:
  - code
  - multilingual
  - legesher
  - transpilation
  - tiny-aya-expedition
  - language-decoded
pretty_name: Language Decoded Data
size_categories:
  - 10K<n<100K
configs:
  - config_name: condition-1-en
    data_files:
      - split: train
        path: data/condition-1-en/train-*.parquet
      - split: validation
        path: data/condition-1-en/validation-*.parquet
  - config_name: condition-2-ur
    data_files:
      - split: train
        path: data/condition-2-ur/train-*.parquet
      - split: validation
        path: data/condition-2-ur/validation-*.parquet
  - config_name: condition-2-zh
    data_files:
      - split: train
        path: data/condition-2-zh/train-*.parquet
      - split: validation
        path: data/condition-2-zh/validation-*.parquet
  - config_name: condition-2-es
    data_files:
      - split: train
        path: data/condition-2-es/train-*.parquet
      - split: validation
        path: data/condition-2-es/validation-*.parquet
dataset_info:
  features:
    - name: code
      dtype: string
    - name: code_en
      dtype: string
    - name: language
      dtype: string
    - name: file_path
      dtype: string
    - name: license
      dtype: string
    - name: token_count
      dtype: int64

Language Decoded | Multilingual Code Dataset

Multilingual Python code datasets for the Language Decoded project (part of Cohere's Tiny Aya Expedition), investigating whether code's reasoning benefit for language models is language-dependent or structure-dependent.

Research Question

Does fine-tuning on non-English code (Python with translated keywords) improve multilingual reasoning as much as English code does?

Prior work (Aryabumi et al., 2024 -- "To Code or Not to Code") demonstrated that including English code in pre-training data improves downstream reasoning performance by approximately 8%. However, that study only tested English code. This dataset enables the natural follow-up: does the reasoning benefit come from the structure of code, or from the language of its keywords?

Dataset Description

This dataset provides filtered, quality-controlled Python source code in four configurations: the original English and three keyword-swapped variants (Chinese, Spanish, Urdu). The source data is drawn from bigcode/the-stack-dedup (Python subset), filtered for quality using the following criteria:

• AST-valid Python only (must parse without errors)
• Permissive licenses only (MIT, Apache-2.0, BSD, etc.)
• 10--1000 lines of code
• Minimum 21 GitHub stars
• No autogenerated files
• SHA-256 deduplication

Keyword-swapped variants are produced using Legesher v0.7.3, which translates Python reserved words (37 keywords, 72 builtins, 66 exceptions) into the target language while preserving code structure and semantics.

Available Configs

Config	Condition	Language	Description
condition-1-en	Condition 1 (control)	English	Unmodified filtered Python from The Stack Dedup
condition-2-ur	Condition 2	Urdu	Keyword-swapped Python -- 37 keywords, 72 builtins, 66 exceptions translated via Legesher v0.7.3
condition-2-zh	Condition 2	Chinese	Keyword-swapped Python -- same transpilation method
condition-2-es	Condition 2	Spanish	Keyword-swapped Python -- same transpilation method

Schema

Column	Type	Description
code	string	Python source code. For condition-2 configs, this is the transpiled (keyword-swapped) version. For condition-1, this is the original English source.
code_en	string	Original English Python source code. Identical to code for condition-1-en.
language	string	ISO 639-1 language code: en, ur, zh, or es.
file_path	string	Original file path in The Stack Dedup.
license	string	SPDX license identifier for the source file.
token_count	int64	Token count computed using the CohereLabs/tiny-aya-base tokenizer.

Experimental Conditions

The Language Decoded experiment uses a ladder of six conditions to isolate the mechanism behind code's reasoning benefit. This dataset currently provides data for conditions 1 and 2:

Condition	Name	Purpose
Baseline	No fine-tuning	Establishes the performance floor
Condition 1	English code	Tests whether code fine-tuning helps at all (replicates Aryabumi et al.)
Condition 2	Keyword-swapped code	Tests whether the language of keywords matters for the reasoning benefit
Conditions 3--6	(planned)	Additional controls not yet included in this dataset

Usage

from datasets import load_dataset

# Load English code (control)
ds = load_dataset("legesher/language-decoded-data", "condition-1-en")

# Load a keyword-swapped variant
ds = load_dataset("legesher/language-decoded-data", "condition-2-ur")
ds = load_dataset("legesher/language-decoded-data", "condition-2-zh")
ds = load_dataset("legesher/language-decoded-data", "condition-2-es")

# Access splits
train = ds["train"]
val = ds["validation"]

Technical Details

Parameter	Value
Source dataset	bigcode/the-stack-dedup (Python subset)
Transpilation tool	Legesher v0.7.3 (legesher-core, legesher-i18n)
Tokenizer	CohereLabs/tiny-aya-base
Base model	CohereLabs/tiny-aya-base (3.35B params)
Train/validation split	90% / 10% (seed 42)
File format	Parquet (snappy compression)
Filtering criteria	AST-valid, permissive licenses, 10--1000 lines, min 21 GitHub stars, no autogenerated files, SHA-256 deduplication

Citation

@misc{language-decoded-2026,
  title={Language Decoded: Investigating Language-Dependent vs. Structure-Dependent Reasoning Benefits of Code},
  author={Madison Edgar and Saad Bazaz and Rafay Mustafa and Sarah Jawaid and Rashik Shahjahan and Khojasteh Mirza and Sohaib Bazaz},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/legesher/language-decoded-data}
}

Links

• Legesher on GitHub
• Tiny Aya Expedition
• bigcode/the-stack-dedup

License

Apache 2.0