Evaluating Large Language Models for Kazakh-Language Educational Assistants: Sentiment Classification and Question Answering with and Without Context

Large language models (LLMs) are increasingly used in education, but evidence on their performance comes mainly from high-resource languages. We evaluated five commercial LLMs in Kazakh in a zero-shot setting: sentiment classification on 300 class-balanced KazSAnDRA reviews and question answering (QA) on the same 300 randomly sampled KazQAD questions with and without the gold supporting passage. Exact McNemar tests with Holm correction found no significant pairwise differences on sentiment classification, one with supporting context, and nine without context under the Containment criterion. For OpenAI and Anthropic, the higher-tier model had a larger advantage over its lower-tier counterpart when the supporting passage was removed (tier-by-regime interaction: OR = 1.92, 95% CI [1.52, 2.43]). Closed-book scores were affected by output format, inflectional morphology, and answer length. Human evaluation of 500 QA responses showed that string-based metrics substantially underestimated closed-book accuracy (0.55 by human judgement versus 0.26 by Containment). Under human judgement, the three higher-scoring models did not differ significantly from one another. Access to a relevant passage narrowed differences between models, although this oracle-context condition did not test a retrieval system. These findings support task-specific evaluation of Kazakh-language educational assistants and the use of answer-matching criteria checked against human judgement.