MayaZ-32M

A small decision model. Input: a state (text or JSON), a question and 2–256 options. Output: a calibrated probability for every option. Three decision types: choice (pick one), noul (yes/no) and score (ordinal).

All options are scored in one encoder pass, side by side, with shared position ids, so option order has no effect (argmax flips under shuffling: 0.03%). Lists above 64 options go through a tournament.

• Backbone: jhu-clsp/ettin-encoder-32m (MIT). 33,877,248 parameters including the decision head.
• Balanced tier and the server default.
• Code, server, browser package and docs: https://github.com/acl-research/MayaZ (MIT).

Results

Metric	Value
Held-out test accuracy (2,485 items)	.781
typed-decisions benchmark (2,000 items)	.732
Calibration-holdout accuracy	.785
ECE after per-bucket temperatures (calibration holdout)	all .033 · 2 .047 · 3–5 .049 · 6–20 .051 · 21+ .069
Prompt-injection / jailbreak detection	0.822 / 0.992

Speed and memory

Measured on the real server under a 10% CPU quota and a 512 MB limit, one ONNX Runtime thread:

Metric	Value
Served file	fp32, 129.5 MB
Latency p50 / p95	1.50 s / 5.31 s
Peak RSS over 200 requests	353 MB
Worst-case request (21 long options, full state), peak RSS in-process	316 MB with the default token budget of 1,024
Browser (ONNX Runtime Web, WASM, 1 thread) p50	~0.3 s

Files

File	Use
mayaz_fp32.onnx	fp32, exact export (served)
mayaz_int8.onnx	weight-only INT8 (MatMulNBits, 8-bit, block 32)
tokenizer.json	Ettin tokenizer, unmodified (MIT)
temperatures.json	Per (type × option-count bucket) temperatures, fitted on the calibration split
config.json	Architecture and renderer settings

The input format (rendering, option markers, truncation) is defined in the GitHub repository: src/mayaz/model/render.py (Python) and web/src/render.js (browser). Use those; do not re-implement it.

How to use

pip install "git+https://github.com/acl-research/MayaZ" huggingface_hub

from huggingface_hub import hf_hub_download
from mayaz.inference import Model

repo = "acl-research/MayaZ-32M"
model = Model(hf_hub_download(repo, "mayaz_fp32.onnx"),
              hf_hub_download(repo, "tokenizer.json"),
              hf_hub_download(repo, "temperatures.json"),
              max_tokens=1024)  # per-pass token budget; keeps peak memory near 316 MB

r = model.decide("choice", "My card was charged twice for the same order.", "What does the customer want?",
                 [("refund", "money back for the duplicate charge"), ("cancel_order", None), ("greeting", None)])
print(r["choice"], r["confidence"], r["probabilities"])

As an HTTP service: docker build -f server/Dockerfile --build-arg MAYAZ_MODEL=mayaz-32m . from the repository, or the Render blueprint in server/render.yaml. In the browser: web/ (ONNX Runtime Web).

Limitations

• Context is 512 positions of state plus the question and options. Longer states are truncated from the end.
• Yes/no questions on paraphrase-style or multi-label tasks tend towards a default answer instead of weighing the evidence.
• On text unrelated to the options the model still picks one, often with moderate confidence. Use the confidence (the server's abstain flag) or add a "none of these" option.
• Typos and slang lower accuracy somewhat without a matching drop in confidence.
• Subjective tasks (emotions, review stars) stay near 0.5 accuracy.
• Not a security control on its own.
• Calibration error in the 3–5 and 21+ buckets is above 0.05. There is no int8-embedding build for this tier.

Training

Soft-label training on decision items rendered from permissively licensed public datasets (train splits only), with soft labels from open-weight teacher models (Apache-2.0):

Dataset	Licence
LocalLLaMA/typed-decisions	Apache-2.0
mteb/banking77	MIT
clinc/clinc_oos (plus)	CC BY 3.0
mteb/amazon_massive_intent (en)	CC BY 4.0
google-research-datasets/go_emotions	Apache-2.0
fancyzhx/dbpedia_14	CC BY-SA 3.0
davidheineman/consumer-finance-complaints-large	Apache-2.0
stanfordnlp/snli	CC BY-SA 4.0
SetFit/amazon_reviews_multi_en	Apache-2.0
google/boolq	CC BY-SA 3.0
nvidia/Aegis-AI-Content-Safety-Dataset-2.0	CC BY 4.0
deepset/prompt-injections (train split)	Apache-2.0
jackhhao/jailbreak-classification	Apache-2.0

Weights: MIT.

Citation

@software{mayaz_2026,
  title        = {MayaZ: small calibrated decision models},
  author       = {Srivastava, Ekansh},
  organization = {ACL Research},
  year         = {2026},
  version      = {1.0.0},
  url          = {https://github.com/acl-research/MayaZ}
}