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

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

This model is a 32 billion parameter GLM-4 variant, fine-tuned by puchuneko, specifically to enhance its understanding and generation capabilities related to Korean culture. It represents a full-weight merge of a pilot QLoRA adapter into the zai-org/GLM-4-32B-0414 base model, making it suitable for direct evaluation in automated harnesses like the K-AI leaderboard.

Key Characteristics & Training:

• Korean Culture Focus: The primary goal of this pilot (v0.1) model is to improve performance on Korean culture-specific subtasks.
• Training Data: Trained on 1,750 rows of data, including culture/knowledge Q&A generated from Korean Wikipedia by Qwen2.5-32B-Instruct, the aya Korean subset, and a 국립국어원 spoken-dialogue subsample.
• Data Decontamination: All training data was rigorously decontaminated against standard Korean benchmarks (KMMLU/CLIcK/HAE-RAE/KoBEST) using a 3-layer n-gram filter to prevent benchmark data leakage.
• Performance Note: While culture subtasks showed a +2.8 percentage point improvement, language subtasks saw a -3 percentage point change, with the aggregate performance remaining within noise levels. The model exhibits some overfitting to quiz-format questions over open-ended generation.

Good For:

• Korean Cultural Q&A: Ideal for applications requiring knowledge and generation related to specific Korean cultural topics.
• Research & Evaluation: Suitable for researchers and developers looking to evaluate the impact of targeted fine-tuning on Korean cultural understanding in LLMs.
• Pilot Projects: Useful for initial explorations into specialized Korean language models, particularly given its pilot (v0.1) status and transparent evaluation.

For full details on the training recipe, data pipeline, and failure log, refer to the GitHub repository.