diff --git a/.gitignore b/.gitignore
index 7e90aa9..f7dfc06 100644
--- a/.gitignore
+++ b/.gitignore
@@ -161,3 +161,4 @@ cython_debug/
 #  and can be added to the global gitignore or merged into this file.  For a more nuclear
 #  option (not recommended) you can uncomment the following to ignore the entire idea folder.
 .idea/
+.vscode/
diff --git a/codepile/ghdiffs/ghdiffs_repo_filter.py b/codepile/ghdiffs/ghdiffs_repo_filter.py
new file mode 100644
index 0000000..0082cfe
--- /dev/null
+++ b/codepile/ghdiffs/ghdiffs_repo_filter.py
@@ -0,0 +1,98 @@
+import re
+from turtle import position
+import pandas as pd
+import requests
+import logging
+from io import StringIO
+import os
+from tqdm import tqdm
+import json
+
+logging.basicConfig(
+    level=logging.INFO,
+    format='%(asctime)s %(levelname)-8s %(message)s',
+    datefmt='%Y-%m-%d %H:%M:%S')
+logger = logging.getLogger(__name__)
+
+
+def read_json_file(json_path: str):
+    with open(json_path, "r") as f:
+        return [json.loads(line) for line in f]
+
+
+def write_json_file(json_path: str, obj: dict):
+    with open(json_path, "w") as jsonFile:
+        json.dump(obj, jsonFile)
+
+
+class GitHubDiffFilter:
+    def __init__(self) -> None:
+        repo_file_url = "https://raw.githubusercontent.com/EleutherAI/github-downloader/master/github_repositories.csv"
+        repo_file_string = StringIO(requests.get(repo_file_url).text)
+        self.single_file_count_threshold = 10_000
+        self.save_every = 1000
+        if not os.path.isdir("tmp"):
+            logger.info(f"Sucessfully added `tmp/` path")
+            os.mkdir("tmp")
+        self.ckpt_path = os.path.join("tmp", "ckpt.json")
+
+        if os.path.exists(self.ckpt_path):
+            self.to_start = read_json_file(self.ckpt_path)["index"]
+        else:
+            self.to_start = 0
+        self.repos_df = pd.read_csv(repo_file_string)
+        self.top_repos_list = self.get_top_repo_list()
+        self.checkpoint_list = []  # Would contain the curr_index of the last saved file.
+        logger.info(f"Size of top repos : {len(self.top_repos_list)}")
+
+    def get_top_repo_list(self) -> list:
+        return self.repos_df.iloc[:, 0].values.tolist()
+
+    def get_diff_path_list(self, github_diff_path: str) -> list:
+        """
+        Get all the files in the given path
+        returns :
+                github_diff_list (list) : GitHub Diff path.
+        """
+        return os.listdir(github_diff_path)
+
+    def __call__(self, github_diff_path, output_file_path):
+        github_diff_files = self.get_diff_path_list(github_diff_path)
+        logger.info(f"Starting to process {len(github_diff_files)} files..")
+        # index and stuffs.
+        content_output_count = 0
+        file_index = 0
+        output_list = []
+        output_file_index = 0
+
+        # Start_ind
+        github_diff_files = github_diff_files[self.to_start:]
+        for github_diff_file in tqdm(github_diff_files, total=len(github_diff_files)):
+            file_index += 1  # File index.
+
+            if file_index == self.save_every:
+                write_json_file(self.ckpt_path, {"index": file_index})
+            github_diff_file_path = os.path.join(github_diff_path, github_diff_file)
+            github_diff_content_list = read_json_file(github_diff_file_path)
+            for ind_content in tqdm(github_diff_content_list, total=len(github_diff_content_list), leave=False):
+                if ind_content["repo_name"] in self.top_repos_list:
+                    output_list.append(ind_content)
+                    content_output_count += 1
+                    self.checkpoint_list.append(file_index)
+
+                if content_output_count == self.single_file_count_threshold:
+                    # If the number of elements in the list is
+                    content_df = pd.DataFrame.from_dict(output_list, orient="columns")
+                    output_inter_file_path = os.path.join(output_file_path, str(output_file_index) + ".json")
+                    content_df.to_json(output_inter_file_path, orient="records", lines=True)
+                    output_file_index += 1
+
+            # Final Chunk goes here..
+            write_json_file(self.ckpt_path, {"index": file_index})
+            content_df = pd.DataFrame.from_dict(output_list, orient="columns")
+            output_final_file_path = os.path.join(output_file_path, str(output_file_index) + ".json")
+            content_df.to_json(output_final_file_path, orient="records", lines=True)
+
+
+if __name__ == "__main__":
+    gh_diff_filter = GitHubDiffFilter()
diff --git a/codepile/ghdiffs/ghdiffs_scrape.py b/codepile/ghdiffs/ghdiffs_scrape.py
new file mode 100644
index 0000000..753446f
--- /dev/null
+++ b/codepile/ghdiffs/ghdiffs_scrape.py
@@ -0,0 +1,183 @@
+import json
+from typing import Any
+import urllib.request
+import argparse
+from pathlib import Path
+import dask.bag as db
+from codepile.dataset import (Dataset, DatasetInfo, RawDataset, Scraper)
+from dask.distributed import Client, progress, LocalCluster
+from unidiff import PatchSet
+
+
+def process_ind_patch(patch_diff) -> dict:
+    """Process patch to get diff data."""
+    patch_parsed_diff: dict = {
+        "hunks": [],
+        "hunks_process": [],
+    }
+
+    patch_parsed_diff["addition_count"] = patch_diff.added
+    patch_parsed_diff["deletion_count"] = patch_diff.removed
+    patch_parsed_diff["src_file"] = patch_diff.source_file
+    patch_parsed_diff["tgt_file"] = patch_diff.target_file
+    if patch_parsed_diff["tgt_file"] == "/dev/null":
+        patch_parsed_diff["file_extension"] = Path(patch_diff.source_file).suffix
+    else:
+        patch_parsed_diff["file_extension"] = Path(patch_diff.target_file).suffix
+    for patch_diff_ind in patch_diff:
+        patch_diff_ind = str(patch_diff_ind)
+        patch_diff_split = patch_diff_ind.split("@@")
+        patch_diff_line = patch_diff_split[2].split("\n")
+        patch_diff_line_numbers = [list(map(int, hunk.strip("-+").split(",")))
+                                   for hunk in patch_diff_split[1].strip().split(" ")]
+        patch_parsed_diff["hunks_process"].append(patch_diff_line_numbers + patch_diff_line[:-1])
+        patch_parsed_diff["hunks"].append(patch_diff_ind)
+    patch_parsed_diff["hunks"] = "".join(patch_parsed_diff["hunks"])
+    return patch_parsed_diff
+
+
+def get_before_file(file_diff: dict, commit_hash: str, repo_name: str, length_threshold: int) -> str:
+    repo_owner, repo_name = repo_name.split("/")
+    if file_diff["src_file"] == "/dev/null":
+        raw_file: Any = ["ADDFILE"]
+    elif file_diff["tgt_file"] == "/dev/null":
+        # If file is deleted, get before file from the raw diff, which will be the full file.
+        raw_file = [line[1:] + "\n" for line in file_diff["hunks_process"][0][3:]]
+    else:
+        # Get raw after file.
+        file_raw_url = (f"https://raw.githubusercontent.com/{repo_owner}/"
+                        f"{repo_name}/{commit_hash}/{file_diff['tgt_file'][2:]}")
+        try:
+            raw_file = urllib.request.urlopen(file_raw_url)
+            raw_file_encoding = raw_file.headers.get_charsets()[0]
+            raw_file = [line.decode(raw_file_encoding) for line in raw_file.readlines()]
+            if length_threshold > 0 and len(raw_file) > length_threshold:
+                return ""
+        except Exception as e:
+            return ""
+        # Iterate over hunks for this file and apply the reverse patch.
+        for hunk in file_diff["hunks_process"]:
+            hunk_list = []
+            for line in hunk[3:]:
+                if line.startswith("-") or line.startswith(" "):
+                    hunk_list.append(line[1:] + "\n")
+            raw_file[hunk[0][0] - 1:hunk[0][0] + hunk[1][1] - 1] = hunk_list
+    del file_diff["hunks_process"]  # Deletes this item from the dict in parent functions
+    return "".join(raw_file)
+
+
+def process_commit(commit_data: dict, config: argparse.Namespace) -> list[dict]:
+    """
+    Process a commit dictionary to get the before files and diff dict.
+
+    Args:
+        commit_data (dict): Dictionary containing commit hash, repo name, and
+        commit message.
+
+    Returns:
+        list[dict]: A list of dicts, where each dict contains the data for a
+        change to a single file.
+    """
+    if config.python_only and commit_data["language_name"] != "Python":
+        return []
+    # Scrape a commit's diff file.
+    diff_url = f"https://github.com/{commit_data['repo_name']}/commit/{commit_data['commit']}.diff"
+    try:
+        diff = urllib.request.urlopen(diff_url)
+        encoding = diff.headers.get_charsets()[0]
+        patch = PatchSet(diff, encoding=encoding)
+        if len(patch) == 0:
+            return []
+    except Exception as e:
+        # print(e, diff_url)
+        return []
+    commit_list: list[dict] = []
+    # Iterate over files within the diff.
+    for patch_ind in patch:
+        if config.ignore_deletions and patch_ind.target_file == "/dev/null":
+            continue
+        if config.diff_length_threshold > 0 and sum(len(hunk) for hunk in patch_ind) > config.diff_length_threshold:
+            continue
+        # Filter non-text files.
+        if patch_ind.added == 0 and patch_ind.removed == 0:
+            continue
+        diff_dict: dict = process_ind_patch(patch_ind)
+        diff_dict["before_file"] = get_before_file(diff_dict, commit_data["commit"], commit_data["repo_name"],
+                                                    length_threshold=config.code_length_threshold)
+        if not diff_dict["before_file"]:
+            # Happens if exception is thrown or file is too long.
+            continue
+        diff_dict["commit"] = commit_data["commit"]
+        diff_dict["message"] = commit_data["message"]
+        diff_dict["repo_name"] = commit_data["repo_name"]
+        diff_dict["language_name"] = commit_data["language_name"]
+        diff_dict["author_name"] = commit_data["author"]["name"]
+        diff_dict["license"] = commit_data["license"]
+        commit_list.append(diff_dict)
+    return commit_list
+
+
+class GitHubDiffDataset(Dataset):
+    def __init__(self, config):
+        self.config = config
+        self.scraper = GitHubDiffScraper(self.config)
+
+    def download(self, *args, **kwargs) -> RawDataset:
+        return self.scraper.scrape()
+
+    def process(self):
+        raise NotImplementedError
+
+    @property
+    def info(self) -> DatasetInfo:
+        return DatasetInfo(
+            id="GitHubDiffDataset",
+            description="Dataset of diffs from GitHub")
+
+    @property
+    def id(self) -> str:
+        return ""
+
+
+class GitHubDiffScraper(Scraper):
+    def __init__(self, config):
+        # TODO: Dask multi-node scheduling here
+        self.config = config
+        cluster = LocalCluster(n_workers=config.n_workers, threads_per_worker=config.threads_per_worker)
+        self.client = Client(cluster)
+        self.read_path = Path(config.read_path)
+        self.save_path = Path(config.save_path)
+
+    def scrape(self) -> RawDataset:
+        meta_spec = {'hunks': str, 'addition_count': int, 'deletion_count': int,
+                     'src_file': str, 'tgt_file': str, 'file_extension': str,
+                     'before_file': str, 'commit': str, 'message': str,
+                     'repo_name': str, 'language_name': str, 'author_name': str,
+                     'license': str}
+        result = (
+            db.read_text(self.read_path).map(json.loads)
+            .map(process_commit, config=self.config).flatten().to_dataframe(meta=meta_spec)
+            .to_parquet(self.save_path)
+        )
+        progress(result)
+        dataset = RawDataset(storage_uris=["https://github.com/CarperAI/Code-Pile"], complete=True)
+        return dataset
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser('codepile dataset tool')
+
+    parser.add_argument('--read_path', type=str)
+    parser.add_argument('--save_path', type=str)
+    parser.add_argument('--n_workers', type=int, default=8)
+    parser.add_argument('--threads_per_worker', type=int, default=1)
+    parser.add_argument('--python_only', type=bool, default=False)
+    parser.add_argument('--diff_length_threshold', type=int, default=1000,
+                        help="Maximum number of lines in the diff for a *single* file. Set to 0 for no limit.")
+    parser.add_argument('--code_length_threshold', type=int, default=1000,
+                        help="Maximum number of lines in code files. Set to 0 for no limit.")
+    parser.add_argument('--ignore_deletions', type=bool, default=True,
+                        help="Ignore file deletion diffs.")
+    config = parser.parse_args()
+    ghdiff_dataset = GitHubDiffDataset(config)
+    ghdiff_dataset.download()
diff --git a/codepile/ghdiffs/test/ghdiffs_dummy.parquet b/codepile/ghdiffs/test/ghdiffs_dummy.parquet
new file mode 100644
index 0000000..e30733b
Binary files /dev/null and b/codepile/ghdiffs/test/ghdiffs_dummy.parquet differ