codetext 0.0.1

Multilingual programming language parsers for the extract from raw source code into multiple levels of pair data

Code-text data toolkit

This repo contains multilingual programming language parsers for the extract from raw source code into multiple levels of pair data (code-text) (e.g., function-level, class-level, inline-level).

Installation

Install dependencies and setup by using install_env.sh

bash -i ./install_env.sh

then activate conda environment named "code-text-env"

conda activate code-text-env

Getting started

Build your language

Auto build tree-sitter into <language>.so located in /tree-sitter/

from src.utils import build_language

language = 'rust'
build_language(language)


# INFO:utils:Not found tree-sitter-rust, attempt clone from github
# Cloning into 'tree-sitter-rust'...
# remote: Enumerating objects: 2835, done. ...
# INFO:utils:Attempt to build Tree-sitter Language for rust and store in .../tree-sitter/rust.so

Language Parser

We supported 8 programming languages, namely Python, Java, JavaScript, Golang, Ruby, PHP, C#, C++ and C.

Setup

from tree_sitter import Parser, Language

raw_code = """
/**
* Sum of 2 number
* @param a int number
* @param b int number
*/
double sum2num(int a, int b) {
    return a + b;
}
"""

parser = Parser()
language = Language("/tree-sitter/cpp.so", 'cpp')
parser.set_language(language)
root_node = parser.parse(bytes(raw_code, 'utf8'))

Get all function nodes inside a specific node, use:

from src.utils.parser import CppParser

function_list = CppParser.get_function_list(root_node)
print(function_list)

# [<Node type=function_definition, start_point=(6, 0), end_point=(8, 1)>]

Get function metadata (e.g. function's name, parameters, (optional) return type)

function = function_list[0]

metadata = CppParser.get_function_metadata(function, raw_code)

# {'identifier': 'sum2num', 'parameters': {'a': 'int', 'b': 'int'}, 'type': 'double'}

Get docstring (documentation) of a function

docstring = CppParser.get_docstring(function, code_sample)

# ['Sum of 2 number \n@param a int number \n@param b int number']

We also provide 2 command for extract class object

class_list = CppParser.get_class_list(root_node)
# and
metadata = CppParser.get_metadata_list(root_node)

Data collection and Preprocessing

The dataset we used to extract was collected by codeparrot. They host the raw dataset in here codeparrot/github-code.

You can create your own dataset using Google Bigquery and the query here

Getting started

For start preprocessing data, define a .yaml file to declare raw data format. (More detail: /data/format/README.md)

python -m src.processing 
<DATASET_PATH>
--save_path <SAVE_PATH>  # path to save dir

--load_from_file  # load from file instead load from dataset cache
--language Python  # or Java, JavaScript, ...
--data_format './data/format/codeparot-format.yaml'  # load raw data format

--n_split 20  # split original dataset into N subset
--n_core -1  # number of multiple processor (default to -1 == using all core)

NOTES: <DATASET_PATH> dir must contains raw data store in .jsonl extension if you pass argument --load_from_file or contains huggingface dataset's