emolero/telelogs_markdown

TeleLogs Dataset (Processed MCQ Format)

This dataset has been extracted from the original netop/TeleLogs dataset and processed into multiple-choice question (MCQ) format for easier evaluation.

Dataset Description

TeleLogs is a telecommunications log analysis benchmark where models must identify the root cause of network issues from 5G wireless network drive-test data and engineering parameters.

Processed Format

This version has been restructured for MCQ evaluation with the following improvements:

• Clean separation of question data, choices, and instructions
• 0-based indexing for answers (0-7 instead of 1-8)
• Extracted choices as a proper array field
• Removed template text from questions

Dataset Structure

Data Instances

Each instance contains:

• question: The actual network data (drive-test logs, engineering parameters, tables)
• choices: Array of 8 possible root causes
• answer: Integer index (0-7) indicating the correct choice

Example:

{
  "question": "Given:\n- The default electronic downtilt value is 255...\n\nUser plane drive test data as follows：\n\n<tables>...",
  "choices": [
    "The serving cell's downtilt angle is too large, causing weak coverage at the far end.",
    "The serving cell's coverage distance exceeds 1km, resulting in over-shooting.",
    ...
  ],
  "answer": 3
}

Data Fields

• question (string): The network data and parameters to analyze. Typically starts with "Given:" and includes:

  • Configuration parameters
  • Drive test data tables
  • Engineering parameters tables

• choices (list of strings): Array of exactly 8 possible root causes

• answer (integer): The correct answer index (0-7), where:

  • 0 = First choice (originally C1)
  • 1 = Second choice (originally C2)
  • ...
  • 7 = Eighth choice (originally C8)

Data Splits

	test
TeleLogs	864

Transformations Applied

This processed version applies three key transformations:

1. Answer Column

• Original: C1, C2, ..., C8
• Processed: 0, 1, ..., 7
• Converted to 0-based indexing to match array positions

2. Choices Column (New)

• Extracted 8 choices cleanly from the original question text
• Each choice stored as array element
• For C8, stops at \n\n separator before actual question data

3. Question Column

• Removed: Template instructions (e.g., "Analyze the 5G wireless network...")
• Removed: Choice listings (C1-C8 with their text)
• Kept: Only the actual question data after \n\n
• Typically starts with "Given:" and includes all data tables

Dataset Creation

Source Data

Original dataset: netop/TeleLogs

Processing Pipeline

1. Load original TeleLogs dataset from HuggingFace
2. Extract test split
3. Parse and separate choices from question text using regex
4. Convert answer format from C1-C8 to 0-7
5. Remove instruction template and choice listings
6. Export to multiple formats (CSV, JSON, Parquet)

Processing Scripts

The processing scripts are available at: Telecom-Bench

Usage

Loading with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("eaguaida/telelogs")

Using with Inspect AI

from inspect_ai import Task
from inspect_ai.dataset import Sample
from inspect_ai.scorer import choice
from inspect_ai.solver import multiple_choice

def telelogs_record_to_sample(record):
    return Sample(
        input=record["question"],
        choices=record["choices"],
        target=chr(65 + record["answer"]),  # Convert 0->A, 1->B, etc.
    )

# Load and evaluate
dataset = load_dataset("eaguaida/telelogs", sample_fields=telelogs_record_to_sample)
task = Task(dataset=dataset, solver=multiple_choice(), scorer=choice())

File Formats

The dataset is available in multiple formats:

• telelogs_test.parquet: Parquet format (recommended, most efficient)
• telelogs_test.json: JSON format (human-readable)
• telelogs_test.csv: CSV format (note: choices stored as JSON string)

Licensing

This dataset maintains the same license as the original TeleLogs dataset.

Citation

If you use this dataset, please cite the original TeleLogs dataset:

@dataset{telelogs2024,
  title={TeleLogs: Telecommunications Log Analysis Dataset},
  author={Original Authors},
  year={2024},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/netop/TeleLogs}
}

Contact

For questions or issues with this processed version, please open an issue at Telecom-Bench.