Skip to content

MLP on Qwen 4B mimicking Jev

X post: a builder reports a small MLP on top of Qwen 4B already reproduces Jev-like decision behaviour.

Discussiondiscussion

This summary is based on how curated lists describe the project; its README was not available to review.

Overview

This is a post on X in which a builder reports that training a small MLP on top of Qwen 4B hidden states already reproduces Jev-like decision behaviour. It belongs to a wider set of community experiments showing that a frozen open model plus a light head can approximate a hosted decision model.

No thread text beyond the list summary was available, so the training data, evaluation and accuracy are unknown. It is mainly a pointer for researchers exploring head-training approaches.

What the decision model decides

TypeQuestionAnswers
ChoicePick one option from Qwen 4B hidden states via a small MLP (inferred)(inferred)—

How it works

  1. Run Qwen 4B over state and question.
  2. Feed hidden states to a small MLP head.
  3. Compare outputs with Jev's decisions.

Strengths

  • Suggests Jev-like behaviour is reachable with a small trained head.

Caveats

  • Social post only; no code, data or metrics in the sources.

At a glance

Type
Discussion
Status
discussion

Use cases and integrations

How curated lists describe it

  • X: builder reports that a small MLP trained on top of Qwen 4B already reproduces Jev-like decision behaviour.

    yibie/awesome-jev

Frequently asked questions

What is claimed?

That a small MLP on Qwen 4B reproduces Jev-like decisions.

Is there code?

None is referenced in the sources.

Are there numbers?

Not in the available summary.

Related projects