processor.py

#!/usr/bin/env python

import json
import argparse
import random
import os
import pandas as pd
import datasets

# Annotation file structure
# {
#     "username": <str>,
#     "fail_reason": {
#         "text": <str>,
#         "vote": <int>
#     },
#     "how_to_fix": {
#         "text": <str>,
#         "vote": <int>
#     },
#     "snippets": [
#         {
#             "text": <str>,
#             "comment": <str>,
#             "file": <str>,
#             "color": <hex>,
#             "vote": 0,
#             "start_index": <int>,
#             "end_index": <int>
#         },
#     ],
#     "id": <str>
# }

SNIPPET_USER_MSG = """Analyse following RPM build log snippet. Describe contents accurately, without speculation or suggestions for resolution.            
Your analysis must be as concise as possible, while keeping relevant information intact.
Snippet: 
{}"""

FULL_CONVERSATION_MSG = """Given following log snippets, their explanation, and nothing else, explain what failure, if any, occurred during build of this package.

  Snippets are in a format of [X] : [Y], where [X] is a log snippet, and [Y] is the explanation.

  Snippets are delimited with '================'.

  Drawing on information from all snippets, provide a concise explanation of the issue and recommend a solution.

  Explanation of the issue, and recommended solution, should take a handful of sentences.

  Snippets:

  {}
"""

ANALYSIS_MSG = "Issue: {issue}\nResolution: {resolution}"


def n_snippets(logs: dict):
    """Count number of snippets in a log."""
    cnt = 0
    for log in logs:
        if logs[log]:
            cnt += len(logs[log]["snippets"])
    return cnt


def none_snippet(logs: dict):
    """Presence of snippet with text set to `None`"""
    for log in logs:
        if logs[log]:
            for snippet in logs[log]["snippets"]:
                if not snippet["text"]:
                    return True
    return False


def total_snippet_annotation_len(logs: dict):
    """Total length of snippet annotations"""
    total = 0
    for log in logs:
        if logs[log]:
            for snippet in logs[log]["snippets"]:
                if snippet["user_comment"]:
                    total += len(snippet["user_comment"])
    return total


def fix_snippets(logs: dict):
    """Set snippet text to equal log content delimited by indices"""
    for log in logs:
        if logs[log]:
            for snippet in logs[log]["snippets"]:
                snippet_text = logs[log]["content"][
                    snippet["start_index"] : snippet["end_index"]
                ]
                if "text" not in snippet or not snippet["text"]:
                    snippet["text"] = snippet_text
    return logs


def cleanup(
    dataset: pd.DataFrame,
    min_how_to_fix_len=10,
    min_fail_reason_len=10,
    min_total_snippet_annot_len=10,
):
    """Cleaning the dataset

    For sake of simplicity we are going to assume following.

    1. Descriptions of issues and fixes shorter than `10` chars are not useful.
    2. Submissions without annotated snippets are not useful.
    3. Submissions with total length of snippet annotations shorter than `10` chars are not useful.
    4. Submissions with snippets set to `None` are not useful."""

    print("Performing dataset cleanup")

    # Fixing snippets with messed up text
    filled_snippets = dataset.logs.apply(fix_snippets)
    dataset["filled_snippets"] = filled_snippets
    dataset.logs = dataset.filled_snippets

    # Setting conditional columns
    dataset["how_to_fix_len"] = dataset.how_to_fix.apply(len)
    dataset["fail_reason_len"] = dataset.fail_reason.apply(len)
    dataset["tot_snippet_annot_len"] = dataset["logs"].apply(
        total_snippet_annotation_len
    )

    # Conditions
    almost_empty_fix = dataset["how_to_fix_len"] < min_how_to_fix_len
    almost_empty_reason = dataset["fail_reason_len"] < min_fail_reason_len
    almost_empty_snippet_annotations = (
        dataset["tot_snippet_annot_len"] < min_total_snippet_annot_len
    )
    none_snippets = dataset["logs"].apply(none_snippet)

    sparse_annotation_criteria = (
        almost_empty_snippet_annotations
        | almost_empty_reason
        | almost_empty_fix
        | none_snippets
    )

    sparse_count = dataset[sparse_annotation_criteria].shape[0]
    print(
        f"Total sparse annotations: {sparse_count} \
        Relative to original dataset: {sparse_count / dataset.shape[0]}"
    )

    # Applying filters
    final_dataset = dataset[~sparse_annotation_criteria]
    final_dataset.reindex()

    return final_dataset


def save_dataset(dataset: pd.DataFrame, path: str):
    """Turn dataframe into parquet"""
    dataset.to_parquet(path)


def load_data(data_path: str):
    """Load json files and return them as list of dicts"""
    data = []
    for root, _, files in os.walk(data_path):
        for file in files:
            if file.endswith(".json"):
                try:
                    with open(os.path.join(root, file), "r", encoding="utf-8") as f:
                        parsed_data = json.load(f)
                        parsed_data["file_name"] = file
                        data.append(parsed_data)
                except json.decoder.JSONDecodeError:
                    print(
                        f"Exception encountered while parsing {os.path.join(root, file)}, skipping."
                    )

    return data


def make_snippet_qa(snippet: dict, user_message_template: str):
    """Create a QnA pair from a snippet"""
    return {
        "question": user_message_template.format(snippet["text"]),
        "answer": snippet["user_comment"],
    }


def make_analysis_qa(
    snippets, issue, resolution, user_message_template, analysis_template
):
    """Create QnA pair from entire annotation."""
    formatted_snippets = "\n=============\n".join(
        [f"[{e['text']}]: [{e['user_comment']}]" for e in snippets]
    )

    return {
        "question": user_message_template.format(formatted_snippets),
        "answer": analysis_template.format(issue=issue, resolution=resolution),
    }


def to_qa_pairs(dataframe: pd.DataFrame):
    """Turn dataframe into list of QnA pairs for training."""
    qa_pairs = []

    for row in dataframe.iterrows():
        sample_snippet_messages = []

        # First element is index
        data = row[1]

        logfiles = data.logs
        # All annoted snippets will be used for final analysis
        annotated_snippets = []
        for file in logfiles:
            if logfiles[file] and "snippets" in logfiles[file]:
                for snippet in logfiles[file]["snippets"]:
                    sample_snippet_messages.append(
                        make_snippet_qa(snippet, SNIPPET_USER_MSG)
                    )
                    annotated_snippets.append(snippet)
        qa_pairs.append(
            make_analysis_qa(
                annotated_snippets,
                data.fail_reason,
                data.how_to_fix,
                FULL_CONVERSATION_MSG,
                ANALYSIS_MSG,
            )
        )
        # Adding individual snippet messages
        qa_pairs.extend(sample_snippet_messages)
    return qa_pairs


def split_dataset(dataset: pd.DataFrame, seed: int):
    """Splits dataset into training and evaluation subset"""
    split = dataset.shape[0] // 5
    dataset = dataset.sample(frac=1.0, random_state=seed)
    train_dataset, test_dataset = dataset[split:], dataset[:split]

    return train_dataset, test_dataset


def main(
    hf_token,
    data_path="results/results",
    sanitize=True,
    seed=42,
    repo_id="fedora-copr/log_detective_qna",
):
    """Process entire directory and turn contents into parquet files."""

    # For reproducibility
    random.seed = seed

    data = load_data(data_path)
    full_dataset = pd.DataFrame.from_records(data)

    if sanitize:
        full_dataset = cleanup(full_dataset)

    # Split dataset
    train_dataset, test_dataset = split_dataset(full_dataset, seed)

    # Format as QnA pairs
    train_dataset, test_dataset = to_qa_pairs(train_dataset), to_qa_pairs(test_dataset)

    dataset = {
        "train": datasets.Dataset.from_list(train_dataset),
        "test": datasets.Dataset.from_list(test_dataset),
    }

    dataset = datasets.DatasetDict(dataset)

    dataset.push_to_hub(repo_id=repo_id, private=True, token=hf_token)

    print("Saving full dataset as parquet...")
    save_dataset(full_dataset, "full_dataset.parquet")


if __name__ == "__main__":
    parser = argparse.ArgumentParser("Dataset Processor")
    parser.add_argument(
        "data_path", type=str, default="results/results", help="Path to annotations."
    )
    parser.add_argument(
        "--sanitize",
        action="store_true",
        help="Run basic data cleanup procedure, dataset build may fail without it.",
    )
    parser.add_argument(
        "--seed",
        type=int,
        default=42,
        help="Seed for random generator to be used when generating split.",
    )
    parser.add_argument(
        "--repo_id",
        type=str,
        default="fedora-copr/log_detective_qna",
        help="ID of Hugging Face repo",
    )
    parser.add_argument(
        "--token", type=str, required=True, help="Token for Hugging face"
    )
    args = parser.parse_args()

    main(
        data_path=args.data_path,
        sanitize=args.sanitize,
        seed=args.seed,
        repo_id=args.repo_id,
        hf_token=args.token,
    )