COMMUNITY OBSERVEDMetric claims: UNVERIFIEDSOURCE AUDITEDDeveloper Tools

SemIf - Direct Choice Probability Extraction from Open Models

Independent research reading choice probabilities directly from open model logits.

Overview & Result

SemIf (formerly OpenJev) by Theo Lee is an independent open-source research initiative exploring how open language models can replicate System One behavior. By extracting output logit probabilities directly over constrained choice sets, SemIf eliminates auto-regressive generation overhead, offering an open-weights counterpart to discrete decision models.

How Jev fits in the loop

  1. Formats decision choices into constrained vocabulary token representations.
  2. Runs a single forward pass through the transformer to obtain next-token logits.
  3. Computes softmax probabilities across the choice tokens to output confidence and rankings.

How to reproduce

  1. Clone https://github.com/TheoLeeCJ/SemIf-OpenJev and install requirements.
  2. Load a supported open-source base model checkpoint (e.g. Qwen, Llama).
  3. Run the choice benchmark script to measure logits directly without multi-token sampling.

Why this build matters

Provides an open-weights foundation allowing developers and researchers to run local, privacy-preserving System One decision models.

Limitations

  • Independent open-source research; not an official TypeSafe API endpoint.
  • Requires GPU hardware to host and run open weights locally.

Patterns