puchuneko/GLM-4-32B-0414-Korean-Culture

GLM-4-32B-0414 Korean-Culture v0.1 (merged full model)

Full-weight merge of the pilot QLoRA adapter puchuneko/GLM-4-32B-0414-Korean-Culture-QLoRA into zai-org/GLM-4-32B-0414, published so automated evaluation harnesses (e.g. the K-AI leaderboard) can run it directly.

This is a pilot (v0.1). See the adapter card for the full honest evaluation: culture subtasks +2.8pp (8/8 flat-or-up), language subtasks −3pp, aggregate within noise, quiz-format overfit on open-ended generation. Full recipe, data pipeline, and failure log: github.com/jwson-automation/korean-glm-merge.

Training data: 1,750 rows — culture/knowledge Q&A generated from Korean Wikipedia by Qwen2.5-32B-Instruct (Apache-2.0), aya Korean subset (Apache-2.0), 국립국어원 spoken-dialogue subsample (weights only; the corpus is not redistributed). All rows decontaminated against KMMLU/CLIcK/HAE-RAE/KoBEST with a 3-layer n-gram filter. No benchmark data was trained on. No SOTA claims.