From Understanding Korean Culture to Screen Control: Everything About Kanana-V Feature Expansion
Kanana-V expands a vision-language model beyond single-image question answering into Korean cultural understanding, document analysis, multi-image reasoning, and GUI interaction. The post details how Kakao built and evaluated these capabilities through large-scale data curation, Korean benchmarks, and training optimizations. Its central conclusion is that language- and task-specific data quality, rather than scale alone, is essential for producing a practical multimodal model. ## Expanding VLM Capabilities - Real-world VLM applications require more than interpreting one image: - Understanding long PDF documents - Comparing multiple images - Interpreting and operating graphical user interfaces - Kanana-V targets these requirements through: - Korean-context understanding - Document and PDF comprehension - Multi-image and long-context processing - GUI grounding for Computer Use Agents (CUAs) - Compared with the similarly sized Qwen3-VL 4B, it achieved broadly comparable results and showed particular strength on Korean-language tasks. ## Curating Korean Interleaved Data - Interleaved datasets alternate images and text, as in blogs, enabling broad knowledge acquisition and stronger in-context learning. - The source collection reached hundreds of terabytes and contained substantial low-quality material, including advertisements, broken images, and duplicated posts. - Kakao used Hugging Face’s Datatrove framework to shard the data and run filtering pipelines in parallel. ## Eight-Stage Data-Cleaning Pipeline - **Image-based document filtering** - Removed broken, tiny, low-resolution, or extreme-aspect-ratio images. - Excluded documents left without valid images. - Used thresholds such as an aspect ratio above 3.0 or dimensions below 28 pixels. - **Language identification** - Applied FastText-based detection. - Retained documents with at least 90% probability of being Korean. - Preserved Korean technical content containing English quotations or code. - **Gopher repetition filtering** - Detected repeated lines, paragraphs, and abnormal 2-gram through 10-gram patterns. - Removed spam and automatically generated advertising content. - **Gopher quality filtering** - Adapted English-oriented rules for Korean. - Lowered the minimum average word length to one character because Korean tokenization often produces short tokens. - Added Korean particles and endings to stopword checks. - **C4 sentence-structure filtering** - Required at least four sentences. - Avoided punctuation-based filtering because Korean writing often omits sentence-final periods. - **FineWeb quality filtering** - Examined short-line ratios, bullet-list frequency, and lines ending in ellipses. - Removed product lists, menus, and similarly unsuitable formats. - **MinHash deduplication** - Used MinHash and locality-sensitive hashing to efficiently identify copied or highly similar documents without performing all pairwise comparisons. - **PII processing** - Masked Korean phone numbers, email addresses, and other personal information. - Cleaned empty text nodes created by image removal and merged adjacent text blocks. ## Impact of Filtering - Approximately 77% of the original data was removed, leaving 23% for training. - Ablation experiments showed that filtered data generally improved performance: - MMVet increased from 33.76 to 36.79. - LLaVA-Wild increased from 75.10 to 78.00. - Korean entity recognition increased from 50.05 to 53.66. - Korean food-menu understanding increased from 44.56 to 47.02. - Korean chart understanding was the exception, declining slightly from 58.33 to 57.43. - The team emphasizes: - Running inexpensive filters before costly ones - Saving intermediate outputs for inspection and reuse - Tuning thresholds for each language - Cleaning related text whenever images are removed from interleaved data The article’s practical recommendation is to treat multimodal model development as an end-to-end data and systems problem: carefully curate culturally relevant data, build language-specific evaluation sets, and optimize training pipelines for each target capability rather than relying solely on larger datasets or models.
Read original(opens in new tab)