Inspect Evals wrap of Thai O-NET M6: filtered multiple-choice items in Thai and English across five school subjects.
unassessed
| Category | knowledge |
|---|---|
| Subcategory | Thai grade-12 national exam (M6) |
| Page status | active |
| Metric | accuracy |
| Direction | higher_is_better |
| Unit | % |
| Dataset size | 397 |
| Dataset licence | Apache-2.0 |
| Publisher | National Institute of Educational Testing Service (exam); OpenThaiGPT / Matichon dump; Inspect Evals packaging by Arthit Suriyawongkul |
onet, as this id, is UK AISI Inspect Evals' task on Thailand's Ordinary National Educational Test for Matthayom 6 (grade 12 / ISCED 3). The model sees a Thai or English stem and lettered choices and must answer with a single letter. Subjects in the dump are Thai language, English, mathematics, science, and social studies. Items that need a written response, cannot be answered from text, or have more than one key are dropped. This is not the ONET slice inside [thai_exam](thai_exam.md), which uses typhoon-ai/thai_exam.
Multiple choice with chain-of-thought. System line: "The question can be in either English or Thai." Inspect multiple_choice(cot=True) asks for ANSWER: $LETTER among A–E. Scorer is choice(). Runnable @task is onet_m6; README command is inspect eval inspect_evals/onet. Configs: default, english, math, science, social, thai. shuffle defaults True. Split is test. Pinned Hub revision 93ffb5e3f3ec630b73e501937805984dd24f2365.
No model card in ModelSpec reports this benchmark yet.