Skip to content
System One

MLP on Qwen 4B mimicking Jev (X)

Builder reports that an MLP trained on top of Qwen 4B works like Jev for typed decisions, and links the weights on Hugging Face as openjev.

Still from MLP on Qwen 4B mimicking Jev (X)Play the video in the post below
Still from x.com

Post on X

Views
108,021
Likes
784
Reposts
48

Read from X on . Counts change daily.

@justALEXWORTEGA trained an MLP (a small feed-forward neural network) on top of Qwen 4B and says it works like Jev. The post links the weights on Hugging Face as openjev, with a demo video.

The current model card describes openjev as Qwen3.5 turned into a cross-encoder that reads a premise and a hypothesis and answers entailment, contradiction or neutral. The 4B v5 checkpoint is the one for typed decisions: give it a state and a closed set of options with their rubrics, and it returns a probability for each, one forward pass per option with nothing generated.

On the public items of JevBench v1.2, the card reports 0.814 overall for the 4B v5 checkpoint against 0.866 for Jev 1.13, with most of the gap on the hard tier (0.622 against 0.730). The card also says the model is not hardened against prompt injection.

Open on X

PrimitivesChoiceScore
Use casesIntent and model routingFeature extraction for machine learning

More like this

Use cases this is tagged with