JIS-124M

JIS-124M is a 124M-parameter decision model for text and JSON. Give it a state and typed questions; it returns candidate probabilities, a selected option, or an expected rubric score without generating text.

• Compact: 124.44M parameters and a 474.7 MiB safetensors checkpoint.
• Typed decisions: supports Choice, Noul, and Score with criteria supplied at inference time.
• Local inference: runs on CPU or CUDA through Hugging Face Transformers.

Model details

Property	JIS-124M
Backbone	GPT-2-derived, affine Pre-RMSNorm, shared candidate scorer
Layers / hidden size	12 / 768
Parameters	124,440,576
Context length	1,024 tokens per state/question/candidate encoding
Vocabulary	50,257 tokens
Release	JIS v3, selected EMA checkpoint

JIS v3 continues JIS v2 on mixed decision datasets, including LocalLLaMA/typed-decisions/train. Per-type calibration temperatures are included in the configuration.

How to use

pip install "transformers==5.17.0" "torch>=2.6" "safetensors>=0.8"

import torch
from transformers import AutoModel

device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModel.from_pretrained("TuanPhat/JIS-124M", trust_remote_code=True).to(device).eval()

questions = {
    "team": {
        "type": "choice",
        "instructions": "Which team should handle this request?",
        "criteria": {
            "billing": "Charges, refunds, and invoices",
            "technical": "Software errors and configuration",
        },
    },
    "refund": {"type": "noul", "instructions": "Is the customer requesting a refund?"},
    "urgency": {
        "type": "score",
        "instructions": "How urgent is this request?",
        "criteria": ["Can wait", "Today", "Blocking the customer"],
    },
}
print(model.system_one("I was charged twice. Please refund the duplicate.", questions))

Type	Criteria	Answer
choice	Candidate IDs mapped to descriptions	choice, probabilities
noul	Optional true / false descriptions	noul: P(true), probabilities
score	Ordered descriptions, lowest first	score: expected level, probabilities, legend

Calls return answers, decisions, and usage. Questions exceeding the context limit return no prediction. JIS scores candidates sequentially and supports text and JSON inputs.

Performance

Typed Decisions, full all/test: 400 cases / 2,000 decisions, scored with the official metrics. Models were measured locally with their native SDKs and released calibration settings.

Model	Parameters	Accuracy ↑	KL ↓	Brier ↓	ECE ↓
JIS-124M	124M	53.70%	0.28328	0.15499	0.07016
Laya	421M	36.35%	0.57207	0.31545	0.17262

Speed and memory

RTX 3060, FP32 parameters, two CPU math threads; warm native calls with five questions per case. Each SDK uses the CUDA autocast precision shown below.

Model	CUDA autocast	Median latency	p95 latency	Peak allocated GPU memory
JIS-124M	FP16	177.7 ms	254.7 ms	512 MiB
Laya	FP16	115.5 ms	152.0 ms	2,389 MiB

Contact

Nguyen Huu Tuan Phat - tuanphatnguyenhuu@gmail.com

Citation

@misc{nguyen2026jis124m,
  author = {Nguyen Huu Tuan Phat},
  title = {JIS-124M: A Compact Typed Decision Model},
  year = {2026},
  url = {https://huggingface.co/TuanPhat/JIS-124M}
}

License

See LICENSE for JIS and GPT-2 code licenses, inherited teacher terms, and dataset attribution.