cassanof/CodeEditSearch

This is a dataset built from CommitPackFT, providing ~1500 commits with diffs for several programming languages:

• Python
• JavaScript
• TypeScript
• Go
• Ruby
• Java
• PHP
• C
• C++
• Rust
• Swift
• Scala
• Bash

The goal of this dataset is to evaluate the ability of models to retrieve a diff given its instruction.

Code To Produce Dataset

Below is the code to reproduce this dataset:

import datasets
from tqdm import tqdm
import difflib


outrepo = "cassanof/CodeEditSearch"
LANGS = ["python", "javascript", "go", "ruby", "java", "php", "c", "c++", "rust", "swift",
         "typescript", "scala", "kotlin", "r", "perl", "haskell", "lua", "shell", "dart", "julia"]

processed = []


def get_udiff(a, b):
    a = a.splitlines()
    b = b.splitlines()
    diff = difflib.unified_diff(a, b, lineterm="")
    return "\n".join(diff)


for lang in tqdm(LANGS):
    print(f"Processing {lang}")
    ds = datasets.load_dataset("bigcode/commitpackft", lang, split="train")
    ds = ds.shuffle(seed=42)
    print(f"{lang}: {len(ds)}")
    ds = ds.filter(lambda x: len(
        x["new_contents"] + x["old_contents"]) < 2500, num_proc=8)
    ds = ds.filter(lambda x: len(x["new_contents"].strip()) > 0 and len(
        x["old_contents"].strip()) > 0, num_proc=8)
    if len(ds) < 2000:
        print(f"Skipping {lang} due to insufficient data")
        continue
    print(f"{lang} after: {len(ds)}")
    ds = ds.select(range(2000))
    diffs = [get_udiff(a, b)
             for a, b in zip(ds["old_contents"], ds["new_contents"])]
    ds = {
        "after": ds["new_contents"],
        "before": ds["old_contents"],
        "diff": diffs,
        "instruction": ds["message"],
        "license": ds["license"],
        "repos": ds["repos"],
        "commit": ds["commit"],
    }
    ds = datasets.Dataset.from_dict(ds)
    ds = ds.filter(lambda x: len(x["diff"].splitlines()) > 10, num_proc=8)
    print(f" ******* Final {lang}: {len(ds)} *******")
    ds.push_to_hub(outrepo, lang)
    processed.append(lang)

print(processed)