• PetteriPano@lemmy.world
    link
    fedilink
    English
    arrow-up
    7
    ·
    2 days ago

    random Chinese character

    It’s beneficial for reasoning to have models trained in a few languages. Chinese is a good one because one character is one word is one token.

    • Meron35@lemmy.world
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 day ago

      Chinese being more token efficient is a myth, and seems to stem from the superficial fact that characters are only visually more space efficient.

      The fact that each Chinese character takes up 3 bytes (as opposed to 1 byte of English), words in Chinese typically require compounds of several characters, and that tokenizers have a limited vocabulary limited to mostly English means that Chinese is actually token inefficient.

      No, Chinese Is Not More Token-Efficient Than English for LLMs | markhuang.ai - https://markhuang.ai/blog/chinese-token-myth

      • elucubra@sopuli.xyz
        link
        fedilink
        English
        arrow-up
        5
        ·
        2 days ago

        When they start hallucinating they output Klingon to me, and to make matters worse, with grammatical errors.

        • urushitan 漆たん@kakera.kintsugi.moe
          link
          fedilink
          English
          arrow-up
          5
          ·
          edit-2
          2 days ago

          Lol I’ve been running linguistic research and it’s funny when they just combine two scripts together into one word

          ● The trigger is identified, and it’s specific.

          agent_1.md:31 src: “Name the trade-off.” → 명取捨之名。 agent_2:32 src: “Name the regime.” → 명regime——名其regime。

          The English imperative “Name the X.” And the corpus renders it correctly 16 other times — 名之 ×11, 名其 ×5, 命名 once.

          명 is the Sino-Korean reading of 名. Same morpheme, wrong script.

          And agent_2:32 is the cleanest evidence I’ve seen for 絡繰’s mechanism: the model wrote 명regime——名其regime — the wrong script and the correct one, eight characters apart, in the same clause. It isn’t ignorant of 名. It produced 名其 immediately after. The meaning resolved correctly both times; the script attribute resolved wrongly the first time and correctly the second.

          That’s exactly what работ法 showed — correct semantics (work), broken script and morphology — and it’s the third confirmed instance of the class, now with a reproducible trigger rather than a one-off.

          It also explains the Russian cases retroactively. document → документ, everything → всё, “correct” → правильно: in each, the meaning landed and the script didn’t. And it predicts why no CJK-native concept ever drifts — there’s no competing script for a morpheme the model only knows in Han.