開放文化基金會（OCF）AI 研究計畫的技術報告總覽。請從下方選擇要閱讀的報告與語言版本。
Technical reports from the Open Culture Foundation (OCF) AI research programme. Pick a report and a language below.
九週實習成果報告：比較兩套獨立建立的評測流程，並以消融實驗拆解 prompt、推論參數、後處理與評分器對分數的影響，最後收斂成一套可重現、可追溯的評測流程。
A nine-week internship report: two independently built evaluation pipelines compared side by side, with ablation studies isolating how prompts, inference parameters, post-processing, and scorers move the numbers — and the reproducible, traceable evaluation workflow that came out of it.
合成影像評測資料集：涵蓋單字、詞彙與句子三個層級共 4,930 張影像，並聚焦 154 個罕見於標準中文的粵文專用字。以兩個 4B 級視覺語言模型（Qwen3-VL、InternVL3.5）實測，指出擴充區 Unicode 字元的辨識失效，並區分視覺辨識不足與語言模型補償兩種效應。
A synthetic image benchmark for Cantonese OCR: 4,930 images across character, vocabulary, and sentence levels, focusing on 154 Cantonese-specific characters that are rare in standard Chinese. Two 4B vision-language models (Qwen3-VL and InternVL3.5) are evaluated, exposing recognition failures on extended-Unicode characters and separating visual recognition deficits from language-model compensation.