Chon-1.0-E4B

Chon-1.0-E4B is a typed-decision model. Given a state (text, JSON or a conversation) and one question with a closed set of options, it returns a calibrated probability for every option:

• choice: pick one of 2-16 labelled options
• yes/no: P(yes)
• score: a distribution over ordered levels

It is a LoRA adapter on google/gemma-4-E4B-it (Gemma 4 E4B, about 4B effective parameters) and serves the TypeSafe /v1/systemone wire format used by JevBench.

Base model	google/gemma-4-E4B-it, unmodified, loaded alongside this adapter
Adapter	LoRA (~558 MB)
Question types	choice (2-16 options), yes/no, score
Input	state up to 6,000 tokens (longer states are truncated, keeping the start); prompt up to 8,192 tokens
Languages	English; Vietnamese also supported
Hardware	one GPU with 24 GB or more for bf16 (measured on one H100 80 GB); on a Mac, Apple silicon with 24 GB or more (--device mps); also runs on CPU with --no-think
Licence	Apache-2.0, as the base model

Use

pip install "chon[serve] @ git+https://github.com/xoai/chon"
python -m chon.serve --model ninhnv/chon-1.0-e4b --port 8000      # downloads the base model and this adapter

On a Mac with Apple silicon and 24 GB or more (uses the Mac's GPU), or on CPU only:

python -m chon.serve --model ninhnv/chon-1.0-e4b --device mps --no-think --port 8000
python -m chon.serve --model ninhnv/chon-1.0-e4b --device cpu --no-think --port 8000

--no-think makes every decision a single forward pass. Use it on a Mac or CPU: the short extra pass for unsure questions generates text, which is slow without a data-centre GPU. Chon-1.0-E2B is the better fit for Macs with less memory.

curl -s localhost:8000/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "Order #4411 shipped on 3 March; the customer reports it arrived damaged on 9 March.",
  "questions": {"decision": {"type": "choice", "instructions": "Route this ticket.",
    "criteria": {"refund": "Refund or replacement", "shipping": "Delivery status", "other": "Anything else"}}}
}'

The response holds answers.decision with choice and probabilities for choice, noul (P(yes)) for yes/no, or probabilities over the levels for score. usage reports input_tokens and output_tokens: most decisions are a single forward pass with no output tokens, while some take a short extra pass whose generated tokens are counted in output_tokens.

What is in this repository

File	What it is
adapter_model.safetensors, adapter_config.json	the LoRA adapter (PEFT)
chon_config.json	serving settings (prompt and calibration) read by the chon server
tokenizer files	the base model's tokenizer and chat template, unchanged
README.md	this card

Evaluation (our measurements, not official board numbers)

set	Intelligence	Calibration (proxy)	Capability (proxy)
held-out replica (720)	72.0	83.8	77.9
JevBench public (231)	77.9	85.7	81.8

Serving, measured on one H100 80 GB, one request at a time: median response 0.08 s, p95 21.4 s; 740 input and 121 output tokens per decision on average.

On CPU only (8 threads of an Intel Xeon Platinum 8480+, bf16, one request at a time, --no-think): median response 3.0 s over 30 short held-out questions (about 230 input tokens; typical questions are about three times longer and take roughly proportionally longer). On a MacBook Pro with 18 GB (Apple M3 Pro) the model did not fit in memory alongside the system; use a Mac with 24 GB or more.

Official JevBench harness (typesafe adapter, public files) against the served package, accuracy: easy 100.0%, original 98.6%, hard 78.4%.

• Held-out set: 720 items in the JevBench v1.5 sealed-set mix (Choice 360 / yes/no 180 / score 180; tiers 14/28/28/30%), written for this project and never trained on.
• Scoring: the board's published formulas where they are public (equal type weights, tier weights 10/20/30/40, yes/no answers inside 0.2-0.8 count as wrong). The board's calibration mapping is not fully published, so the calibration column is our proxy.

Licence

Apache-2.0, as google/gemma-4-E4B-it.