dataset_split

Take Home Test: Reformat a Public Dataset for LLM Training

Delivery

• Reformatted Dataset: a.json
• Transformation Code: data_split.py
• Statistics: 106155
• Performance Metrics: 1.0415241718292236 seconds

---

Test content

Take Home Test: Reformat a Public Dataset for LLM Training

Objective

The goal of this task is to prepare public datasets for more effective use in training and fine-tuning Large Language Models (LLMs). You are required to reformat a specific subset of a public dataset into a structured, consistent format to facilitate its usability.

Detailed Instructions

1. Dataset Selection and Preparation

• Dataset: You are assigned the Headline subset of the AdaptLLM/finance-tasks dataset.

• Task Description: Each entry in the input column contains multiple "Yes" or "No" questions alongside their respective answers. Your task is to:

  • Develop a Python script to parse and separate each question and its answer from the entry.

  • Save each question-answer pair in a structured JSON format as follows:

    {
      "id": "<unique_identifier>",
      "Question": "<question_text>",
      "Answer": "<answer_text>"
    }

  • You are encouraged to introduce additional attributes if needed to preserve the integrity and completeness of the information. Adding relevant tag information is strongly recommended.

• Automation Requirement: The task must be completed using Python. Manual editing or data manipulation is strictly prohibited. Your script should efficiently handle variations in data format within the column.

2. Deliverables

• Reformatted Dataset: Provide the schema of the final format you adopted for saving the results.
• Transformation Code: Submit the complete code used for converting the dataset into the designated format.
• Statistics: Report the total number of question-answer pairs extracted from the dataset.
• Performance Metrics: Document the time taken to complete the dataset cleanup and transformation process.