Naluri-561M-encoder

Instinctive, multilingual decisions in one pass.

Naluri (Indonesian for instinct; Neural Answering via Label Understanding and Rapid Inference) is a small System One decision model. This is the 561M member of the family and the strongest so far, built on XLM-RoBERTa large. You give it a text or JSON state and a typed question with its options written out. It returns calibrated probabilities over those options in a single forward pass, in about 14 ms per question on a laptop GPU. It never generates text.

Model	Backbone	Parameters	System One Bench	Latency (GPU, p50)
Naluri-279M-encoder	XLM-RoBERTa base	278.6M	50.4 (`main`, M6-pairs); 58.0 (`m9` branch)	4.9 ms
Naluri-308M-encoder	mmBERT-base	307.5M	61.0	9.2 ms
Naluri-561M-encoder (this model)	XLM-RoBERTa large	560.9M	64.9	13.8 ms

It supports the same three question types as Jev and Laya:

Type	Options	Returns
`noul`	yes / no (optional descriptions)	p(true)
`choice`	2+ named options (optional descriptions)	one probability per option
`score`	ordered levels	one probability per level; expected level = score

Code, training pipeline and full report: github.com/jerichosiahaya/naluri.

Load with the bundled naluri.py (see Usage). The standard transformers pipeline and the hosted inference widget can't run Naluri: its decision head is custom code.

Usage

pip install torch "transformers>=4.48" safetensors huggingface_hub

from huggingface_hub import hf_hub_download
import importlib.util

# naluri.py (input format + model + loader) ships with this repo
spec = importlib.util.spec_from_file_location("naluri", hf_hub_download("jerichosiahaya/Naluri-561M-encoder", "naluri.py"))
naluri = importlib.util.module_from_spec(spec); spec.loader.exec_module(naluri)

tok, model = naluri.from_pretrained("jerichosiahaya/Naluri-561M-encoder", device="cpu")   # or device="cuda"

state = {"message": "Pesanan saya sudah seminggu belum sampai dan CS tidak membalas."}
questions = {"complaint": {"type": "noul", "instructions": "Is `message` a complaint?"},
             "urgency": {"type": "score", "instructions": "How urgent is `message`?",
                         "criteria": ["not urgent", "somewhat urgent", "very urgent"]}}
print(naluri.predict(tok, model, state, questions, device="cpu"))

Named fields work. Put the inputs in named fields and point at them with backticks in the question.
Option descriptions matter. Naluri reads them, so clear, distinct descriptions give better answers.
Limits: the question plus options must fit in about 192 tokens, and the full input is capped at 512 tokens.
Uncertainty is meaningful: probabilities near 0.5 (or several options close together) are the cases to send to a human or an LLM.

Results: System One Bench

System One Bench: 16 suites, 3,250 cases, run with the bench's own run_decisions.py against serve.py and scored with score.py at 91f4f28. Index = mean suite accuracy. All models ran on the same laptop GPU (RTX 5050). 0 failed cases.

	279M (M9)	308M	561M
Index	58.0	61.0	64.9
Calibration error (ECE)	0.027	0.057	0.024
Accuracy on its 50% most confident answers	67.8%	70.6%	77.7%
Latency per question (p50)	4.9 ms	9.2 ms	13.8 ms

Suite	279M (M9)	308M	561M
XNLI en / fr / ar	62.7 / 62.7 / 54.7	76.0 / 68.0 / 56.7	71.3 / 67.3 / 65.3
Emotion	70.0	84.5	80.5
MASSIVE intent (ar)	52.5	65.5	76.5
MASSIVE scenario (ar)	59.5	63.5	59.0
typed-decisions	46.9	54.6	56.7
MIRACL pairwise (ar)	69.0	76.0	82.0
Arabic offensive	75.5	79.5	77.0
AG News	80.0	79.5	82.5
Arabic sentiment	60.0	59.0	63.0
MIRACL rerank (ar)	34.0	34.0	39.0
Banking77	29.0	27.0	36.5
BoolQ	60.0	55.0	56.5
SST-2	74.5	68.5	81.0
SST-5	36.5	28.0	44.5

Best overall and best calibrated member of the family. Sentiment, Arabic intent and retrieval improve the most.
Many options are still the weakest area (Banking77, MIRACL rerank), limited by the 192-token option budget. Narrow the options first, then let Naluri decide among the top 12 or fewer.
Scores are from our own run; the bench maintainers have not run this model yet.

Training

Input format (the same layout as Laya): <s> {type} question: {instructions} </s> <mask> option 1 <mask> option 2 … </s> {state} </s>. Each option is scored from the encoder vector at its own <mask> by a small MLP head, and a softmax over a question's options gives the answer distribution.
Objective: cross-entropy against one-hot labels or a teacher's soft probabilities, followed by one calibration temperature per question type (choice 1.53, score 1.33, noul 2.19).
Recipe: one run from the XLM-RoBERTa large checkpoint on the M9 data (125k typed questions), the same data and sampling as the 308M model: up to 5 epochs with early stopping (patience 1), encoder learning rate 2e-5, head 5e-4, task-balanced sampling (p ∝ √size), synthetic decisions weighted 2× and minimal pairs 3×. The token embeddings (256M parameters) were frozen so training fits an 8 GB GPU; 305M parameters were trained. Epoch 4 was kept, picked on the mean of per-task dev accuracies (SIB-200 and XCOPA validation splits plus synthetic dev sets): 0.697, vs 0.674 for the 308M model. About 6.5 hours of compute on an 8 GB RTX 5050 laptop GPU.
Field layout: 50–70% of the public-task rows are shown as named fields (premise, hypothesis, passage, question…), so the model reads /v1/systemone-style states.
Augmentation: options shuffled, labels renamed, instructions paraphrased, varying option counts, and noul, choice and score mixed.
Weights: stored in fp16 (the encoder) and loaded in fp32. The largest probability difference against the fp32 training checkpoint was 0.0017, with the same top answer on all 85 questions (65 benchmark cases, fp32 inference).

Training data

About 125k typed questions. The licenses below are as listed on each dataset's card; check them yourself before any use.

Dataset	Use	License (per dataset card)
XNLI	entailment (noul / choice)	CC BY-NC 4.0
MASSIVE (`mteb/amazon_massive_intent`)	intent (choice / noul)	CC BY 4.0 (original) / Apache 2.0 (mirror)
Amazon Reviews Multi (mirror)	review stars (score) + polarity	Amazon's own terms (non-commercial)
TextDetox multilingual toxicity	toxicity	OpenRAIL++
papluca/language-identification	language ID	derived from Amazon Reviews, XNLI and STSb
dair-ai/emotion	emotion	other
RACE	reading comprehension, plus yes/no option checks	non-commercial research
SQuAD 2.0	is the question answerable from the passage? (noul)	CC BY-SA 4.0
MultiRC (SuperGLUE)	is this answer correct? (noul)	see the dataset card
UCI SMS Spam	spam	unknown on the Hub
MasakhaNEWS, Yahoo Answers Topics, DBpedia-14	topic classification	see each dataset card
Synthetic decisions, ops decisions and minimal pairs (20k)	business decisions in 17 languages	written by DeepSeek V4 Flash, labeled with probabilities by DeepSeek V4 Pro

The SIB-200 and XCOPA validation splits were used only to pick the epoch. No test split of any eval set was trained on.

License

These weights are released under CC BY-NC 4.0 (non-commercial), because several training datasets restrict commercial use. The base model (XLM-RoBERTa large) is MIT. The code on GitHub is Apache 2.0. For commercial use, retrain on data you are licensed to use; the pipeline is in the GitHub repo.

Limitations

Accuracy is below the best LLM-based decision models. The System One Bench leaders (decoders with 4B–27B parameters) score 76–83. Use Naluri as a fast first pass and send low-confidence cases to a larger model.
BoolQ stays near chance (56.5%) and MASSIVE scenario (ar) is lower than in the 308M model.
Many options (more than about 12) are weak.
Bigger and slower than the smaller members: 1.1 GB fp16, about 14 ms per question on a laptop GPU.
Unrelated fields can sway yes/no answers (measured on the 308M model, same data). Put only the fields a question needs into the state.
Seeds: this is a single run.