diff --git a/data_analysis/pr_code_reviews/create_sample.py b/data_analysis/pr_code_reviews/create_sample.py
new file mode 100644
index 0000000..68c4de2
--- /dev/null
+++ b/data_analysis/pr_code_reviews/create_sample.py
@@ -0,0 +1,109 @@
+from tqdm import tqdm
+import pandas as pd
+from datasets import load_dataset, Dataset
+
+
+source = "https://huggingface.co/datasets/bigcode-data/pr_code_reviews/resolve/main/groupby_pull_request_id"
+
+
+def process_bucket(bucket_id="00"):
+    """Build a dataset with fours cols: pull_request_info, head_repo_info, base_repo_info and events (list of dicts/events) with only main info"""
+    pr_00 = load_dataset(
+        "parquet",
+        data_files=[f"{source}/{bucket_id}.parquet"],
+        split="train",
+    )
+    df_00 = pr_00.to_pandas()
+    print(
+        f"number of PRs in bucket {bucket_id}: {len(df_00['pull_request.id'].unique())}"
+    )
+
+    # group by pull_request.id
+    grouped_data = df_00.groupby("pull_request.id")
+
+    # we dropped some columns that didn't seem useful
+    pull_request_info_cols = [
+        "repo.name",
+        "public",
+        "pull_request.id",
+        "pull_request.number",
+        "pull_request.title",
+        "pull_request.body",
+        "pull_request.state",
+        "pull_request.user.login",
+        "pull_request.user.id",
+        "pull_request.created_at",
+        "pull_request.closed_at",
+        "pull_request.merged_at",
+    ]
+    head_info_cols = [
+        "pull_request.head.label",
+        "pull_request.head.ref",
+        "pull_request.head.user.login",
+        "pull_request.head.user.type",
+        "pull_request.head.repo.owner.login",
+        "pull_request.head.repo.owner.type",
+        "pull_request.head.repo.license.name",
+        "pull_request.head.sha",
+    ]
+    base_info_cols = [
+        "pull_request.base.label",
+        "pull_request.base.ref",
+        "pull_request.base.sha",
+        "pull_request.base.user.login",
+        "pull_request.base.user.type",
+        "pull_request.base.repo.owner.login",
+        "pull_request.base.repo.owner.type",
+        "pull_request.base.repo.license.name",
+        "pull_request.base.repo.default_branch",
+        "pull_request.base.repo.description",
+        "pull_request.base.repo.language",
+        "pull_request.base.repo.watchers_count",
+        "pull_request.base.repo.open_issues_count",
+        "pull_request.base.repo.forks_count",
+    ]
+    pull_request_cols = pull_request_info_cols + head_info_cols + base_info_cols
+    # drop "repo.name", "repo.id", "public" so they are not duplicated and keep relevant columns that might change
+    event_cols = [
+        col
+        for col in df_00.columns
+        if (not col.startswith("pull_request."))
+        and col not in ["repo.name", "repo.id", "public"]
+    ] + [
+        "pull_request.head.label",
+        "pull_request.head.ref",
+        "pull_request.head.sha",
+        "pull_request.title",
+    ]
+
+    pr_dict = []
+    for name, group in tqdm(grouped_data):
+        group_dict = {}
+        # sort events by date created_at
+        group = group.sort_values(by=["created_at"])
+        events = group.to_dict(orient="records")
+        # we get info from the first event
+        group_dict["pull_request_info"] = group[pull_request_info_cols].to_dict(
+            orient="records"
+        )[0]
+        group_dict["head_repo_info"] = group[head_info_cols].to_dict(orient="records")[
+            0
+        ]
+        group_dict["base_repo_info"] = group[base_info_cols].to_dict(orient="records")[
+            0
+        ]
+        group_dict["events"] = group[event_cols].to_dict(orient="records")
+        pr_dict.append(group_dict)
+    return pr_dict
+
+
+# process 3 random buckets
+pr_dict_00 = process_bucket("031")
+pr_dict_01 = process_bucket("093")
+pr_dict_02 = process_bucket("02B")
+pr_dict_03 = process_bucket("005")
+
+buckets = pr_dict_00 + pr_dict_01 + pr_dict_02 + pr_dict_03
+hf_dataset = Dataset.from_pandas(pd.DataFrame(data=buckets))
+print(hf_dataset)
+hf_dataset.push_to_hub("pr_code_reviews_sample")