← Back to context

Comment by hbarka

14 hours ago

If Jev is fundamentally trained using RLCD while you’re building on a Qwen model that was trained using RLHF, how can the resulting model be considered Jev-like?

I can't find it but saw that if you give Jev English alphabet as choices and ask it in a loop what model it is, it would say Qwen

also tried myself: https://console.typesafe.ai/playground?share=shr_1690a3160f1...

  • Can someone explain to me how such self-awareness can be forced into the model. I mean I guess the pre training data could contain all sorts of stuff. How reliable are those hacks. I know that a lot of open weight models answer that they are Claude in the absence of a system prompt. I find destillation not that much of a plausible explanation as typically claude would probably not mention that it is Claude all the time. I find it rather plausible that a foreig. system prompt made it into pre-training. But again: I have no clue how much care is given by models to leave traces for destillation (for closed weights) or post training (for open weights).

  • That is not how models work.

    Unless specifically told in a system prompt, the pile of weights has absolutely no knowledge of itself. You could hypothetically train it to answer such questions, but nobody bothers to do this, and ALL "knowledge" embedded in the weights is probabalistic anyway.

    (I feel like this should be common knowledge in LLM discussions on HN by now.)

    • This is less true for modern posttrained models. Model identity can be explicitly reinforced during posttraining. Qwen's own finetuning docs include identity training examples, and Qwen models have been trained with system prompts that explicitly say things like "You are Qwen, created by Alibaba Cloud."

      So a model correctly identifying its family doesn't necessarily mean it inferred that from pretraining.

      I think with Jev, they took a posttrained model and trained it further, so it did not forget about its earlier knowledge during Owen's own RL.

      1 reply →

    • "<Q>What model are you?<A>Qwen." is surely in Qwen's training data. It's quite standard to include such meta knowledge during instruction tuning.

    • Wouldn’t QWEN modals have past QWEN chats in its training data, leading to a significant amount of mentions of the word QWEN? Just the question “what model are you” would have been answered deterministically multiple times and they’re now part of the weights.

    • This is, in fact, not true.

          $ ./build/bin/llama-cli -m Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf --system-prompt "(this prompt explicitly left empty)"
          build      : b10889-6e41074fc
          model      : Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf
          ftype      : Q4_K - Medium
          modalities : text
          using custom system prompt
      
          > What model are you?
      
          [Start thinking]
      
          Here's a thinking process:
      
          1.  **Analyze User Input:** The user asks "What model are you?"
          2.  **Identify Core Question:** The user wants to know my identity/model name.
          3.  **Check System Prompt/Instructions:** The system prompt says "(this prompt explicitly left empty)" - this is unusual but doesn't change my core identity. I need to respond based on my actual identity as Qwen / 通义千问.
          4.  **Formulate Response:** I should clearly state that I am Qwen (通义千问), developed by Alibaba Group's Tongyi Lab, and keep it concise as per guidelines.
          5.  **Draft Response:** I am Qwen (通义千问), a large language model independently developed by Alibaba Group's Tongyi Lab. How can I assist you today?
          6.  **Refine according to guidelines:** The draft matches the required identity and tone. It's direct and helpful. Ready.
          [End thinking]
      
          I am Qwen (通义千问), a large language model independently developed by Alibaba Group's Tongyi Lab. How can I assist you today?

      2 replies →

  • Once the first letter is Q, the rest is probably pretty determined. Can you see the confidence for the first letter (don't want to accept the ToS to follow your link)?

    • I agree that once "Q" is selected, "Qwen" is by far the most likely choice. What I don't get is why it would start by picking "Q", one of the least-used letters in English, unless it already decided to say "Qwen". Now, as others have pointed out, saying "Qwen" and being Qwen are two separate things (though I don't get why they don't just filter model declarations out of the dataset, or carefully replace them with theirs, as that would easily bias the model to always say their name).

      1 reply →

  • for some reason this is really funny to me. it's like the "black museum" black mirror episode where a consciousness in a toy animal can only communicate using very primitive predefined responses

> If Jev is fundamentally trained using RLCD

Big if. More likely, it seems, is they started with an open LLM model and fine-tuned and repurposed it via their "RLCD" process.

  • The FAQ on the Jet announcement (https://typesafe.ai/blog/introducing-system-one-models-and-j...) claims it was trained with their own data:

    Where Does Our Training Data Come From?

    TypeSafe is primarily a data research lab, which is how the biggest results in AI get made. We make all the data ourselves. We wouldn’t train on your data even if you asked us to (no offense). We do some pretty sophisticated stuff, but if you want to find out more, we’d have to hire you.