Multi-Domain Instruction and QA (1k Samples)

---

License: CC-BY-SA-3.0
Language: English
Tags: question-answering, instruction-following, multi-domain, reasoning, mathematics, medical, career-guidance

---

Dataset Summary

This dataset is a curated collection of 1,000 instruction-following and question-answering samples, designed to evaluate the robustness and generalizability of language models across diverse domains. It aggregates data from four distinct, high-quality sources, covering mathematical reasoning, medical knowledge, general instruction following, and career guidance.

This dataset was compiled for the StageRAG Project to serve as a practical benchmark for multi-domain RAG systems and language model evaluations.

Supported Tasks

• question-answering: The primary task for this dataset is evaluating a model's ability to provide accurate answers to questions from various knowledge domains.
• instruction-following: A significant portion of the dataset (from Dolly-15k) tests a model's ability to follow general, freeform instructions.

How to Use

The dataset can be easily loaded using the Hugging Face datasets library:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("YourUsername/YourDatasetName")

# The dataset contains a single 'train' split
print(dataset['train'])

# Expected output:
# {
#   'source': 'gsm8k', 
#   'prompt': 'Natalia sold 48 liters of milk in the morning and 34 liters in the afternoon. She had 12 liters of milk left. How many liters of milk did she have at the beginning?', 
#   'response': 'Natalia sold 48 + 34 = 82 liters of milk. She had 12 liters of milk left. So, at the beginning, she had 82 + 12 = 94 liters of milk.'
# }

Dataset Structure

Data Fields

Each sample in the dataset contains the following fields:

• source (string): The identifier of the original dataset the sample was taken from (e.g., gsm8k, medqa, dolly-15k, career-guidance).
• prompt (string): The user question, instruction, or problem statement. The original column names (question, instruction, etc.) have been standardized to prompt.
• response (string): The ground-truth answer or response. The original column names (answer, response, etc.) have been standardized to response.

Data Splits

The dataset consists of a single split: train, containing all 1,000 samples. No predefined validation or test splits are provided.

Dataset Creation

Curation Rationale

The primary motivation for this dataset was to create a small but highly diverse benchmark. The chosen sources cover a spectrum of tasks required by modern LLMs:

• Logical & Mathematical Reasoning (gsm8k)
• Specialized Domain Knowledge (MedQA)
• Open-Ended General Instruction Following (dolly-15k)
• Practical, Real-World Q&A (career-guidance-qa-dataset)

By combining these, the dataset can be used to quickly assess a model's performance beyond a single domain.

Source Data

A total of 250 samples were randomly selected from each of the following four datasets:

⚠️ WARNING: The Pradeep016/career-guidance-qa-dataset does not have an explicit license. Before using this combined dataset for any purpose beyond personal experimentation, you should replace these samples with data from a permissively licensed source.

Annotations and Preprocessing

The creation process involved minimal preprocessing. The key steps were:

1. Randomly sampling 250 rows from each source dataset.
2. Standardizing the column names to prompt and response for consistency.
3. Adding a source column to maintain provenance for each sample.
4. Combining the four sets of samples into a single file.

Limitations and Bias

• Small Size: With only 1,000 samples, this dataset is intended for quick evaluation and benchmarking, not for training or fine-tuning models.
• No Test Split: The dataset does not have a dedicated test split, so users must create their own validation/test splits if needed.
• Inherited Biases: Any biases present in the original source datasets are inherited in this collection.
• English Only: The dataset is entirely in English.

Citation

If you use this dataset in your work, please cite the original sources. You can also cite this dataset compilation as:

@dataset{Darrencxl_Multi_Domain_Instruction_QA_2025,
  author = {Darren Chai Xin Lun},
  title = {Multi-Domain Instruction and QA},
  year = {2025},
  publisher = {Hugging Face},
  version = {1.0.0},
  url = {https://huggingface.co/datasets/darren0301/domain-mix-qa-1k}
}

License

This derivative dataset is licensed under the Creative Commons Attribution-ShareAlike 3.0 International (CC BY-SA 3.0).

This license was chosen to comply with the "ShareAlike" requirement of the databricks/databricks-dolly-15k dataset, which is one of its sources. Under this license, you are free to share and adapt the dataset for any purpose, even commercially, as long as you give appropriate credit and distribute any derivative works under the same license.