Qwen on VibePapa — Tech from the Boseong River https://boseonggang.com/tags/qwen/ Recent content in Qwen on VibePapa — Tech from the Boseong River Hugo en-us Fri, 23 Oct 2026 07:06:59 +0900 Six Old Books Translated for $1.20: Why I Stopped Cutting and Scanning https://boseonggang.com/posts/six-books-translated-for-1-dollar-20-cents-deepseek-vs-local-qwen/ Fri, 23 Oct 2026 07:06:59 +0900 https://boseonggang.com/posts/six-books-translated-for-1-dollar-20-cents-deepseek-vs-local-qwen/ I planned to cut the spines off six Korean commentary volumes and scan them. Then I found the English originals in a public-domain library and translated all six with DeepSeek for about $1.20. Why local Qwen was not good enough, what the pipeline checks, and what is still unreviewed. A Bible Library That Answers With Its Sources: Sermons, a Commentary and Three Bibles in One Search https://boseonggang.com/posts/bible-library-that-answers-with-its-sources/ Fri, 23 Oct 2026 05:48:40 +0900 https://boseonggang.com/posts/bible-library-that-answers-with-its-sources/ I put hundreds of sermon transcripts, a scanned commentary and three Bible texts into one local search that answers with numbered sources. Here is what works, the bug that said 'no data', the tunnel that kept dropping, and why re-reading a scanned book is still unsolved. Local Qwen vs Gemini for Fixing a Transcript: 50 Minutes Free vs 2.3 Minutes for 9 Cents https://boseonggang.com/posts/local-qwen-vs-gemini-correcting-a-sermon-transcript/ Thu, 22 Oct 2026 08:31:33 +0900 https://boseonggang.com/posts/local-qwen-vs-gemini-correcting-a-sermon-transcript/ I wanted to correct speech-to-text errors in a Korean sermon transcript without sending it to a cloud API, so I tried a local Qwen model. It took about 50 minutes and made new mistakes. Gemini took 2.3 minutes and cost $0.086. The comparison, with numbers and what each model got wrong. I Marked Who Was Speaking by Ear, Then Graded the AI's Speaker Labels Against My Marks https://boseonggang.com/posts/grading-ai-speaker-labels-against-my-own-ears/ Thu, 22 Oct 2026 07:50:08 +0900 https://boseonggang.com/posts/grading-ai-speaker-labels-against-my-own-ears/ My local speech-to-text pipeline labeled who was talking in a 39-minute audio file, and I wasn't sure it was right. I marked one speaker's turns by hand while listening, then compared. The AI was right 91.5% of the time, and a script did the math instead of a language model.