JevAlt
Three 4B models choose between options and return probabilities, running locally via Jev API.
Tested on Linux
Use it
Linux x86_64
You need first
- Runtimes:
python>=3.11 - A GPU is optional
- At least 3 GB of RAM
install jevalt[serve] from git (gguf extra fails without gcc)
pip install "jevalt[serve] @ git+https://github.com/mertkayacs/jevalt"
Check that it works
jevalt --helpThe output should contain “usage: jevalt”.
First thing to try
cat > /tmp/check_f1.py << 'PY'
from jevalt.engine import DecisionEngine
class DummyBackend:
name="dummy"
def count_tokens(self,t): return len(t.split())
def field_logits(self,text,compiled): return [[5.0 - i*2 for i,_ in enumerate(f.symbols)] for f in compiled.fields]
def continue_reasoning(self,prefix,max_tokens): return "trace"
b=DummyBackend()
e=DecisionEngine(b, model="dummy")
payload={"state":"I was charged twice for March. Please refund the duplicate.","questions":{"team":{"type":"choice","instructions":"Which team should handle this ticket?","criteria":{"billing":"payments, refunds","technical":"bugs, outages","sales":"prices, contracts"}}}}
res=e.predict(payload)
print(res)
assert "probabilities" in res["answers"]["team"] and "choice" in res["answers"]["team"]
print("F1_OK probabilities choice")
PYpython3 /tmp/check_f1.pyWhat it does
- Choose between options and get probabilitiesTested
How to use it
- In Python, create a
DecisionEnginefromjevalt.enginewith a backend and model name. - Call
predictwith astatetext and aquestionsentry of typechoice, withinstructionsandcriteriafor each option. - Read
choiceandprobabilitiesunderanswersin the result.
- In Python, create a
- Get calibrated probabilities for each question typeTested
How to use it
- In the
questionspassed topredict, settypetochoice,scoreornoul. - For
score, givecriteriaas a list of levels. Fornoul, give a single criterion. - Read the
probabilities,scoreornoulvalue for each question underanswers.
- In the
- Get an unknown option with reasoning tracesTested
How to use it
- Add
"abstain": Trueand"reasoning": "on"to the request passed topredict. - Read
unknownandreasoningfor each question underanswersin the result.
- Add
How far pickfit got
Linux x86_64
- Installs
- Starts
- 3 of 3 features work
- Works offline
Good for
- route support tickets to the handling team
- check inputs against a written policy
- classify incident severity
An alternative to Kev-4B and Laya.
How pickfit tested it
Each run starts from a clean machine and follows the steps above. How pickfit verifies
Linux x86_64
Its main function works
- Replayed from scratch
- Works offline
- Checked on 1 Nov 2026
- Version 0.1.0
Differences the test machine hit
- README: pip install "jevalt[serve,gguf] @ git+https://github.com/mertkayacs/jevalt"
In the test: pip install "jevalt[serve] @ git+..."; llama-cpp-python wheel fails to build: CMake Error Could not find compiler gcc, no binary wheel available for this platform