phorknorak/khmer-asag-benchmark

Khmer Automatic Short-Answer Grading Benchmark

Dataset summary

This dataset supports research on automatic short-answer grading in Khmer. It contains 1,184 answers graded by one teacher and written by 203 students from two Cambodian high schools, one in a province and one in a city. The corpus covers 41 questions from Biology, History, Geography, and Earth Science.

The repository provides two configurations:

• full: all 1,184 complete answers;
• curated-no10c: 909 answers after removing an inconsistently graded Biology subset.

The dataset was developed by Phork Norak, an undergraduate student, for the undergraduate report Explainable AI for Automated Grading of Khmer Language Short-Answer Questions in Education, under the supervision of Dr. Khim Chamroeun. As a student research project, it may have limitations or errors. Constructive feedback and corrections are welcome.

Intended uses

• Research on Khmer automatic short-answer grading.
• Comparison of classical, neural, Transformer, and language-model approaches.
• Research on explainable educational NLP and word attribution.
• Reproducibility studies using the documented random split and seed.

The dataset is not intended to support fully automated high-stakes grading without teacher review.

Data fields

Field	Description
SchoolID	Public non-identifying school code (SCH001 or SCH002)
ClassID	Class code retained for educational analysis
Subject	School subject
StudentID	Original pseudonymous student code (for example, S004)
QuestionID	Question identifier
Question	Question text
Reference	Teacher reference answer
Answer	Student answer
Student Score	Score assigned by the teacher
Max Score	Maximum possible score for the question
Year	Collection year

The five-class ordinal label used by the benchmark is derived as round(4 * Student Score / Max Score) and clipped to the range 0 to 4.

Dataset creation

The answers come from real semester examinations. The curated-no10c configuration removes one Biology subset after the grading pattern was found to be inconsistent with the rest of the corpus. No synthetic answers are included.

Personal and sensitive information

Direct student and school names are not included. School identifiers were replaced with public non-identifying codes. The original pseudonymous student codes and class codes are retained for educational analysis, but no mapping to real identities is released. An automated review found no email addresses, URLs, phone-number patterns, or obvious self-identification phrases in the text. Responsible users should not attempt to re-identify or contact students, teachers, or schools.

Limitations and bias

• The corpus contains data from only two schools and four subjects.
• All scores come from one teacher, so they measure agreement with that teacher rather than universal correctness.
• The score distribution and question coverage are not balanced.
• Performance may not transfer to other schools, subjects, grade levels, or forms of Khmer.
• The curated configuration reflects a research decision to remove one noisy subset.

Licence

The dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International Licence (CC BY-NC 4.0). Attribution is required, and commercial use is not permitted. See the included DATA_LICENSE.md and the official licence. The MIT licence of the associated software does not apply to this dataset.

Responsible use

Users should not attempt to re-identify or contact students, teachers, or schools. The dataset should not be used as the sole basis for high-stakes educational decisions. These responsible-use guidelines describe the intended ethical use of the data and do not add conditions to the CC BY-NC 4.0 licence.

Citation

@software{phork_khmer_asag_2026,
  author = {Phork, Norak},
  title = {Explainable AI for Automated Grading of Khmer Language Short-Answer Questions in Education},
  year = {2026},
  url = {https://github.com/PhorkNorak/Explainable-AI-for-Automated-Grading-of-Khmer-Language-Short-Answer-Questions-in-Education}
}