ENPT

simple-jev

Last updated:

RuntimeApache-2.0

⚠️ Why not a model

Turns any compatible open LM into a typed classifier via next-token logits — no fine-tuning required.

Technical specs

Base LLMs
any-open-lm
Decision types
choice, score, noul
Features
demo-api, rfdt, model-agnostic
License
Apache-2.0

Links

simple-jev is Featherless AI’s open stack that turns any compatible open LLM into a typed classifier via next-token logit extraction.

How it works

simple-jev reads decision probabilities from the next-token distribution:

  1. Format the decision question as a prompt
  2. Extract logits for each option token
  3. Convert to probabilities for choice, score, or noul

It does not require fine-tuning — any open LLM that follows instructions can be used as the backend.

Why it’s not a model

simple-jev is purely an inference-time technique:

  • No model training or fine-tuning
  • Works with any compatible open LLM
  • The “decision” capabilities come from prompt engineering + logit extraction

Key features

  • Model-agnostic: Use any open LLM as the backend
  • Demo API: Try it at simple-jev.featherless.ai
  • RFDT support: Integrates with the Featherless ecosystem

Caveats

  • Probabilities are not calibrated
  • Performance depends entirely on the base LLM
  • Does not reproduce TypeSafe’s RLCD training method