Thread

1. hbarka · 1789982542 · focus · HN ↗

   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?

   1. mohsen1 · 1789986029 · focus · HN ↗

      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: <a href="https:&#x2F;&#x2F;console.typesafe.ai&#x2F;playground?share=shr_1690a3160f19c2e4851b6a2883bcf17a790" rel="nofollow">https:&#x2F;&#x2F;console.typesafe.ai&#x2F;playground?share=shr_1690a3160f1...

      1. bityard · 1789993110 · focus · HN ↗

         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 &quot;knowledge&quot; embedded in the weights is probabalistic anyway.

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

         1. mohsen1 · 1789994475 · focus · HN ↗

            This is less true for modern posttrained models. Model identity can be explicitly reinforced during posttraining. Qwen&#x27;s own finetuning docs include identity training examples, and Qwen models have been trained with system prompts that explicitly say things like &quot;You are Qwen, created by Alibaba Cloud.&quot;

            So a model correctly identifying its family doesn&#x27;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&#x27;s own RL.

            1. janalsncm · 1790019222 · focus · HN ↗

               Right, if a model says it is Qwen there is no way to distinguish a ModernBert fine tuned with Qwen completion data from a Qwen model fine tuned with completion data.

               It’s also entirely possible that they used completions from a pool of open weight models.

Unofficial Hacker News client; not affiliated with Y Combinator.