Dataflex-selection Dataset (Alpaca-format)

This dataset contains two splits:

• train: Alpaca-format data for instruction tuning from Openhermes_10w  
• validation: multiple-choice reasoning data for evaluation, which includes not only the ground-truth answers but also model-generated predictions  

Train Dataset (Alpaca Format)

A high-quality instruction–response dataset in Alpaca format, designed for training and fine-tuning large language models on instruction-following tasks.   The dataset includes structured triplets:

• instruction: user instruction  
• input: optional additional input  
• output: target answer generated by the model  

---

Validation Dataset

The validation split is derived from the MMLU benchmark and has been further enhanced with GPT-5–generated Chain-of-Thought (CoT) reasoning. Each sample contains the full question, the correct answer, and a model-generated prediction, enabling detailed evaluation and error analysis.

This split can be used for automated evaluation, benchmarking, data selection experiments (e.g., TSDS / NEAR), and studying reasoning failure cases.

Each validation sample contains the following fields:

• instruction: the full question text (including the Question and answer options)  
• input: an empty string, kept for Alpaca-format consistency  
• output: the ground-truth answer option text (e.g., "A. ...")  
• prediction: the model’s predicted answer or its generated reasoning  
• answer: the final correct answer label (e.g., "A")