Datasets:

cs-552-2026-databand
/

general_knowledge_dataset

license: other
task_categories:
  - question-answering
  - text-classification
language:
  - en
pretty_name: General Knowledge SFT Dataset
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
      - split: valid
        path: data/valid.jsonl

General Knowledge SFT Dataset

This dataset contains the exact train and validation data used for the general knowledge LoRA SFT model in the MNLP project Specialize and Merge: Post Training Qwen3-1.7B for Multi Skill Reasoning.

The dataset has two splits.

Split	Rows	Purpose
train	26,120	LoRA SFT training split
valid	2,000	LoRA SFT validation split

Sources

The SFT data was built from six multiple-choice educational and science-oriented sources.

Source	Train	Valid	Main coverage
EduAdapt	388	17	Educational multiple-choice questions
EduQG	3,200	121	School and science question generation data
Kaggle LLM Science	6,342	223	Science exam questions
NCERT_MCQs	435	15	Textbook-style school MCQs
OpenBookQA	4,697	532	Elementary science and open-book reasoning
SciQ	11,058	1,092	Crowdsourced science questions

Preprocessing

All sources were converted into the same multiple-choice schema: a question, a list of answer choices, and one correct answer. We removed examples that could not be converted reliably, such as rows with missing questions, missing choices, missing answer keys, answer keys that did not point to any choice, or too few valid choices. We also removed duplicate examples. For sources where the same distractor text appeared more than once, we kept one copy of the repeated choice.

To match the TA benchmark interface, which can use option labels up to A to T, we randomly relabeled the answer choices with labels from A to T while preserving the correct answer text. We then constructed the final splits by uniformly sampling examples for each correct label. The train split contains exactly 1,306 examples for each correct label from A to T. The valid split contains exactly 100 examples for each correct label from A to T.

Format

Each row contains:

Column	Description
prompt	User-facing multiple-choice prompt
answer	Correct answer letter
source	Original source dataset
split	train or valid
messages_json	JSON string containing the chat messages used for SFT
meta_json	JSON string containing auxiliary metadata

The SFT user message follows this pattern:

Question text

Choices:
D. option text
N. option text
C. option text
O. option text

The assistant target is always exactly one boxed letter, for example:

\boxed{D}

During training, completion-only SFT was used. The system message, user prompt, question, choices, and chat-template tokens were masked. Cross-entropy loss was applied only to the final assistant boxed answer.

System prompt

The general knowledge specialist used a fixed system prompt instructing the model to choose exactly one option and return only one boxed letter, such as \boxed{A}, \boxed{C}, \boxed{J}, or \boxed{T}, with no explanation or extra text.

Related benchmark dataset

The held-out model-selection benchmarks are stored separately in:

cs-552-2026-databand/general_knowledge_benchmark

They were not used for SFT training.

Dataset citations and source references

Kaggle LLM Science Exam: Kaggle. LLM Science Exam. https://www.kaggle.com/competitions/kaggle-llm-science-exam

EduQG: nlztrk. EduQG Dataset, LLM Science Exam Format, 34k. Kaggle dataset. https://www.kaggle.com/datasets/nlztrk/eduqg-dataset-llm-science-exam-format-34k

EduAdapt: EduAdapt multiple-choice educational question data, as used in the project preprocessing pipeline. https://huggingface.co/datasets/notefill/eduadapt

NCERT_MCQs: NCERT textbook-derived multiple-choice questions, as used in the project preprocessing pipeline. https://huggingface.co/datasets/goenkalokesh/NCERT_MCQs/blob/main/NCERT_MCQs.csv

OpenBookQA: Mihaylov, T., Clark, P., Khot, T., and Sabharwal, A. 2018. Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering. EMNLP 2018. https://allenai.org/data/open-book-qa

SciQ: Welbl, J., Liu, N. F., and Gardner, M. 2017. Crowdsourcing Multiple Choice Science Questions. Workshop on Noisy User-generated Text. https://allenai.org/data/sciq

Intended use

This dataset is intended for reproducibility of the MNLP general knowledge specialist. It is suitable for multiple-choice answer selection experiments where the model must return a boxed option letter.

Limitations

This dataset is built from public educational and science-oriented multiple-choice sources, so its coverage is strongest for school-level science, textbook-style facts, and commonsense questions. It is not a complete factual knowledge dataset, and it does not guarantee reliable performance on specialized professional domains such as medicine, law, or finance. Because the examples come from different original sources, their licensing and annotation quality may vary. The dataset should therefore be used for reproducible model training and evaluation, not as an expert system or a high-stakes assessment tool.