sarathi-balakrishnan/browser-decision-model-base

SB Decision Model (base)

What it is. A 151M-parameter decision model by Sarathi Balakrishnan, trained, calibrated, compressed to 159 MB and packaged to run inside a browser tab. Give it a text and your own list of options, each written as name: when to choose it, and it returns a calibrated probability per option in one pass. Nothing is retrained per use case: the option definitions are the policy. A reserved ask option is trained to win when the deciding fact is absent from the text, so irreversible actions can be routed to a person instead of guessed.

What it is not. Not an LLM: it never generates text. No multi-step reasoning, arithmetic, counting, or world knowledge beyond what the text says. English only. It is the router / guardrail / triage layer in front of an LLM, not a replacement for one.

How it runs. In the browser tab: downloaded once and cached, GPU when the browser has one, CPU otherwise, inference off the main thread, zero network requests after load. The browser build is verified against the original model on every release. The same weights serve Node, Python, and an HTTP endpoint through one API.

Get it

Python	pip install sb-decision · https://pypi.org/project/sb-decision/
JavaScript, React, Node	npm install sb-decision · https://www.npmjs.com/package/sb-decision
Weights	this repository, public: SBDecision.load() fetches them with no token
Author	https://huggingface.co/sarathi-balakrishnan

from sb_decision import SBDecision

m = SBDecision.load()  # sarathi-balakrishnan/browser-decision-model-base, no token needed
print(m.decide("Package arrived broken, second time this month. I want a manager.",
               ["refund: customer wants money back for a damaged, wrong, or missing item",
                "replace: customer wants the product again",
                "escalate: angry, threatens to leave, or asks for a manager"]))
# {'decision': 'escalate', 'band': 'auto', 'scores': {'escalate': 0.90, 'ask': 0.09, 'refund': 0.01, 'replace': 0.01}}

import { SBDecision } from "sb-decision";

const m = await SBDecision.load();   // browser (cached after the first load) or Node, no token needed
const r = await m.decide("Package arrived broken, second time this month. I want a manager.", [
  "refund: customer wants money back for a damaged, wrong, or missing item",
  "replace: customer wants the product again",
  "escalate: angry, threatens to leave, or asks for a manager",
]);
// { decision: "escalate", band: "auto", scores: { escalate: 0.90, ask: 0.09, refund: 0.01, replace: 0.01 }, ms: 24 }

band is auto at or above 0.7, review at or above 0.4, otherwise abstain; if ask wins it is always abstain. Both thresholds are arguments, not model constants. Independent tags: mode="multi". Many texts in one pass: scores_many / scoresMany.

Results

split	n	top1_agreement	auroc_macro	ece	brier	ask_recall	false_ask_rate
val	6738	0.826	0.979	0.035	0.034	0.725	0.010
heldout	4745	0.765	0.947	0.030	0.054	0.752	0.024

Per held-out domain:

domain	n	top1_agreement	auroc_macro	ece	brier	ask_recall	false_ask_rate
airline-disruptions	952	0.844	0.974	0.032	0.040	0.709	0.006
open-source-issue-triage	936	0.745	0.953	0.035	0.057	0.829	0.034
real-estate-lead-routing	950	0.730	0.925	0.066	0.060	0.619	0.017
restaurant-reservations	952	0.747	0.953	0.070	0.059	0.798	0.028
veterinary-intake	955	0.759	0.935	0.043	0.057	0.808	0.036

Held-out domains (never in training or validation): veterinary-intake, real-estate-lead-routing, airline-disruptions, open-source-issue-triage, restaurant-reservations. Re-measured by an independent validation pass.

Latency on an M3 laptop: 38 ms per short decision in the browser (88 ms at 512 tokens), 80 ms p50 on the HTTP endpoint for an 80-word request.

Targets not met by this release, reported rather than hidden: held-out top-1 0.765 against a 0.80 target (the reference labels agree with themselves 0.883 of the time, so 0.765 is 87% of what the metric can measure); option-order invariance max |dp| 0.23 against 0.01 (mean 0.015). Met: calibration error 0.030 against 0.05, ask recall 0.752 against 0.75, false ask 0.024 against 0.10.

Variants

model	params	held-out top-1	where it runs
SB Decision Model base	151M	0.765	browser and server
SB Decision Model large	399M	0.796	server

Same input format and API; the large model is the one to use when download size does not matter.

Known weaknesses

• Decides from the text, not the world; cannot look anything up.
• Weak at math, counting, date arithmetic, multi-hop conditions.
• Options are matched by the meaning of the definition; a vague definition gives a vague score. Describe what the text says, not what the agent should do.
• ask means "not enough information to choose between these options". It does not check whether the chosen action has its parameters: "Send the updated contract to the client" is a confident send even with no client named. Keep the option named ask with its definition "not enough information to decide".
• Single-choice mode is overconfident when two options genuinely apply; use independent mode when options are not mutually exclusive.
• English only in this release.

Data

No real people, companies, emails, or phone numbers appear in the training data.

License

Apache-2.0. See NOTICE.

Citation

@software{sbdm2026, author={Balakrishnan, Sarathi}, title={SB Decision Model: a decision model with user-supplied options that runs in the browser}, year={2026}, url={https://huggingface.co/sarathi-balakrishnan/browser-decision-model-base}}