えー! そんじゃ最初にcpt読み込ませてる意味が分からん・・追加教育出来ないの? ってトコでclaudeのchat長制限来たw
Always include ALL your source files in the files list when training. If you want to add new files later:
# Include both old AND new files files = ["yin.wav", "yang.wav", "new1.wav", "new2.wav"]
This explains perfectly why your model keeps generating the same patterns - it's literally forgotten everything except the last training session! You need to retrain with all files together to get the mixed, varied output you want.
続きをchatgptのほうに説明してもらったら
A Different File With the Same Name
The model assumes it's seeing the same "concept"
→ But now the content is different (e.g. yin.wav was soft before, now it’s harsh)
→ The model says “oh! I must have misunderstood this keyword”
→ It adapts to fit the new version, and forgets the old one
This is catastrophic forgetting — classic in small models.
小さなモデルに生じがちなカタストロフィック忘却に陥るんだと?
「だから、さっきそう言っただろ!」的にキレないとこが有り難いね。多分以前に指摘されてる解決法なのに、その時点では何を言われてるのか不明だったからな