LocalLLaMA/typed-decisions

Typed Decisions

A benchmark for typed probabilistic decisions. A model gets one piece of unstructured state and answers five typed questions about it at once, and every answer is a probability distribution, not a single label.

The schema follows the System One primitives (noul, choice, score) used by TypeSafe AI, so a row replays against any API with that shape. The benchmark is independent: it is not affiliated with TypeSafe and does not reproduce their Jev model.

Leaderboard

test split: 400 cases, 2,000 decisions. This table lists general models scored zero-shot: they have never seen these workflows or their question schemas. Models that were fitted or fine-tuned on train are in the second table below; the two tables are not comparable.

#	Model	Kind	Accuracy ↑	KL from gold ↓	Brier ↓	ECE ↓	p50 latency	Price / 1M input
1	Quyet-1.0-Large (31B) via Decision Models (quyet-1-large) (#38)	general, zero-shot, open weights	0.791	0.281	0.114	0.093	3,335 ms	$0.035
2	meraGPT Decider 1 (sd-1)	general, zero-shot	0.768	0.096	0.052	0.180	526 ms	$0.03
3	Baseten Mammoth 1 (27B, from Qwen3.5-27B) via OpenRouter (baseten/mammoth-1:free)	general, zero-shot	0.768	0.122	0.063	0.153	415 ms	free tier
4	Amos 1 medium (9B) (#31)	general, zero-shot, open weights	0.7645	0.151	0.080	0.103	118 ms (H100)	open weights
5	Amos 1 large (27B) (#31)	general, zero-shot, open weights	0.761	0.126	0.069	0.097	357 ms (H100)	open weights
6	pplx-decider-v1.1-27b (27B) (#37)	general, zero-shot, open weights	0.753	0.441	0.160	0.060	–	open weights
7	d3 (27B) (#37)	general, zero-shot, open weights	0.749	0.409	0.146	0.049	–	open weights
8	Baseten Elephant 1 (9B, from Qwen3.5-9B) via OpenRouter (baseten/elephant-1:free)	general, zero-shot	0.748	0.134	0.069	0.162	352 ms	free tier
9	Darwin-27B-ZTC (27B) (#17)	general, zero-shot, open weights	0.743	0.204	0.097	–	–	open weights
10	d3-mini (4B) (#37)	general, zero-shot, open weights	0.743	0.319	0.123	0.047	–	open weights
11	Liquid AI d1 (d1:free)	general, zero-shot	0.742	0.475	0.155	0.124	525 ms	free tier
12	linne-oak 1.0 (31B) (#35)	general, zero-shot, open weights	0.739	0.709	0.191	–	1.3 s (8 concurrent)	open weights
13	vLLM-SR d3 (26B, FP8) via Decision Models (s1-pro) (#38)	general, zero-shot, open weights	0.739	0.437	0.151	0.124	3,405 ms	$0.08
14	d3-turbo (26B) (#37)	general, zero-shot, open weights	0.736	0.438	0.162	0.069	–	open weights
15	clef (27B) (#37)	general, zero-shot, open weights	0.728	0.191	0.103	0.025	–	open weights
16	openjev (27B) (#37)	general, zero-shot, open weights	0.728	0.506	0.191	0.101	–	open weights
17	TypeSafe Jev 1.13.0	general, zero-shot	0.727	1.442	0.148	0.144	710 ms	$0.042
18	Blink v0.3 (26B-A4B, NVFP4) via Decision Models (s1-fast) (#38)	general, zero-shot, open weights	0.727	0.153	0.079	0.129	3,274 ms	$0.019
19	d3-flash (9B) (#37)	general, zero-shot, open weights	0.717	0.322	0.131	0.031	–	open weights
20	Featherless Simple Jev (Qwen3.6-35B-A3B-classifier)	general, zero-shot	0.716	0.488	0.176	–	–	free demo
21	jev-vision-27b (Qwen3.8-27B + LoRA)	general, zero-shot, open weights	0.714	0.318	0.129	0.026	199 ms (H100)	open weights
22	clef-flash (9B) (#37)	general, zero-shot, open weights	0.710	0.209	0.111	0.016	–	open weights
23	prima-ratio + 12B	general, zero-shot	0.702	0.564	0.234	0.146	700 ms	local GGUF
24	JEV-35B (35B MoE) (#37)	general, zero-shot, open weights	0.700	0.526	0.169	0.062	–	open weights
25	JEV-27B-VL (28B) (#37)	general, zero-shot, open weights	0.688	0.544	0.194	0.113	–	open weights
26	clef-omni (35B) (#37)	general, zero-shot, open weights	0.673	0.247	0.123	0.030	–	open weights
27	OpenDecider-small	general, zero-shot, open weights	0.671	0.211	0.117	–	40 ms	open weights
28	d1-3B (3B) (#37)	general, zero-shot, open weights	0.656	0.286	0.144	0.037	–	open weights
29	d3-nano (2B) (#37)	general, zero-shot, open weights	0.629	0.545	0.211	0.113	–	open weights
30	Bongard-mini	general, zero-shot, open weights	0.594	0.256	0.132	0.067	225 ms	open weights
31	Jeff-Gemma4-E2B	general, zero-shot, open weights	0.561	0.403	0.219	0.188	2,272 ms	open weights
32	d3-lite (0.8B) (#37)	general, zero-shot, open weights	0.531	0.518	0.244	0.130	–	open weights
33	GEV-26B-Decide (26B) (#37)	general, zero-shot, open weights	0.522	0.423	0.215	0.066	–	open weights
34	Jeff-Qwen3.5-2B	general, zero-shot, open weights	0.511	0.460	0.237	0.203	1,346 ms	open weights
35	JEV-27B (27B) (#37)	general, zero-shot, open weights	0.495	0.550	0.253	0.136	–	open weights
36	Jeff-Qwen3.5-0.8B	general, zero-shot, open weights	0.483	0.679	0.313	0.251	662 ms	open weights
37	d3-edge (0.6B) (#37)	general, zero-shot, open weights	0.425	1.176	0.456	0.341	–	open weights
38	THX-01 (0.3B) (#37)	general, zero-shot, open weights	0.377	0.620	0.320	0.203	–	open weights
–	Prior (ignores the input)	reference	0.470	0.347	0.189	0.088	–	–
–	Uniform (same probability on every option)	reference	0.308	0.444	0.238	0.169	–	–

Fitted or fine-tuned on train

These models were trained on this benchmark's train split, so they are not zero-shot and their scores are not comparable with the table above.

#	Model	Kind	Accuracy ↑	KL from gold ↓	Brier ↓	ECE ↓	Reported latency
1	Decider-4B (Qwen3.5-4B + LoRA) (#36)	general, then fine-tuned on train and other decision data, temperature-calibrated	0.8205	0.301	0.132	0.031	117 ms (A800)
2	Arbiter-4B (from NeoHorse-Jev-4B + LoRA) (#34)	general, then fine-tuned on train and other decision data, temperature-calibrated	0.817	0.092	0.048	0.184	133 ms (A800)
3	Darwin-27B-ZTC-v5 (27B, full fine-tune) (#33)	general, then fine-tuned on train	0.812	–	–	–	–
4	ministral3-3b-reasoning-typed-decisions (Ministral 3 3B Reasoning, full fine-tune) (#30)	specialist	0.808	0.077	0.042	0.182	373 ms (GB10)
5	Amos 1 large (27B), per-question priors fitted on train (#31)	general, then fitted on train (one log-prior bias per question and option; no weights changed)	0.8025	0.122	–	–	357 ms
6	OpenDecider-large-td (Qwen3-Next-80B-A3B + LoRA)	general, then fine-tuned on train	0.801	0.081	0.044	–	440 ms
7	od1-typed-decisions (Qwen3.5-4B)	specialist	0.797	0.082	0.045	0.156	–
8	OpenDecider-nano (400M encoder)	general, then fine-tuned on train	0.796	0.079	0.043	–	17 ms
9	OpenDecider-small-td (Qwen3-4B + LoRA)	general, then fine-tuned on train	0.792	0.080	0.043	–	40 ms
10	Phocinae-Largha-150M-v1 (mmBERT-small, 144.3M)	specialist	0.791	–	–	–	21.0 ms (RTX 5090)
11	OpenDecider-medium-td (Qwen3-30B-A3B + LoRA)	general, then fine-tuned on train	0.788	0.081	0.044	–	214 ms
12	soft-decider-421m	specialist	0.774	–	–	0.141	50 ms
13	Qwen3.5-4B-typed-decisions (Qwen3.5-4B + LoRA, Clef head)	general, then fine-tuned on train	0.7725	0.0992	0.0543	0.139	–
14	embeddinggemma-2-ft-typed-decision (0.3B) (#37)	fine-tuned on train	0.770	0.119	0.065	0.155	–
15	Laya typed-decisions	fitted on train	0.766	–	–	–	–
16	gemma3-270m-typed-decisions (Gemma 3 270M, full fine-tune)	specialist	0.761	0.132	0.072	0.148	68 ms
17	SafeDecide-140M (mmBERT-small, 140M) (#39)	general, then fine-tuned on train	0.754	0.126	0.069	0.147	80 ms (1 CPU thread)
18	ProKope-421M-8K (421M, 8K context)	specialist	0.7185	–	0.385	–	43 ms
19	gemma3-270m-typed-decisions (Gemma 3 270M, full fine-tune)	general, fine-tuned on train	0.702	0.156	0.091	0.128	68 ms
20	Bekko System One v0 400M, scored by us	trained on train and other data	0.668	0.214	0.113	0.143	808 ms
21	vega-08b-public-intents (0.8B) (#37)	specialist (bundled typed-decisions adapter active)	0.663	0.435	0.191	0.074	–
22	ModernBERT-base (149M), ours	specialist, fitted per workflow	0.646	0.223	0.119	0.179	349 ms
23	MiniLM-L6 (22M), ours	specialist, fitted per workflow	0.587	0.262	0.143	0.108	22 ms
24	Bekko System One v0 68M, scored by us	trained on train and other data	0.537	0.293	0.160	0.116	176 ms
25	Bekko System One v0 17M, scored by us	trained on train and other data	0.483	0.344	0.203	0.136	33 ms

Quyet-1.0-Large, served by Decision Models, has the highest accuracy among zero-shot models (0.791). meraGPT Decider 1 is second on accuracy (0.768) and has the lowest KL (0.096) and Brier (0.052) of any zero-shot model. It leads Jev on every question type (noul 0.840 vs 0.775, choice 0.733 vs 0.720, score 0.739 vs 0.696), its distributions sit far closer to the gold (KL 0.096 vs 1.442), it is faster end to end, and it costs less per token. It answers POST /v1/systemone at meragpt.com, so the typesafe-sdk works against it by setting TYPESAFE_BASE_URL.

To add a model, score it on test with the full distributions and open a discussion with the numbers and the mode (specialist or general) it used.

Benchmarking a model with HF Jobs

One command scores any model that serves vLLM's /v1/systemone (any model without a native server works this way; serve it with a current nightly). It runs one request per case (state + all five questions) over the test split and prints accuracy, KL from gold, Brier and ECE:

hf jobs run --flavor a100-large --timeout 2h \\
  --env MODEL_ID=your-org/your-model \\
  --secrets HF_TOKEN \\
  --volume hf://datasets/LocalLLaMA/typed-decisions:/repo:ro \\
  pytorch/pytorch:2.9.1-cuda12.8-cudnn9-devel \\
  bash /repo/bench/score_with_jobs.sh

• Pick the --flavor for the model's size: l4x1 up to ~2B, l40sx1 or a100-large for ~27B, h200 above.
• bench/score_with_jobs.sh installs vLLM nightly, serves the model, scores all 400 cases, and writes results.json.
• The run finishes by creating the eval YAML and opening a PR on the model repo with the four scores, so they show on the Hub leaderboard once merged. Add --env MODE=specialist if you fitted the model on train.

Who scored each row

• Scored by us: meraGPT Decider 1, TypeSafe Jev 1.13.0, Liquid AI d1, Baseten Mammoth 1 and Elephant 1, Featherless Simple Jev, the three Jeff models, the three rows served by Decision Models (#38), the Bekko and Adaptive Classifier rows, and the Prior and Uniform references.
• Scored by Hugging Face: the rows linked to #37, run on Hugging Face Jobs.
• Reported by the model's creator: every other row, in its own discussion.

Reading a score

Gold is the mean of three samples from a teacher of roughly 4B-class capability, so a score measures agreement with that teacher, not correctness. A better model can score lower wherever the teacher is wrong (it missed a duplicate invoice whose ID matched an earlier one).

Reference	Accuracy	What it is
Prior	0.470	the floor: label frequencies, ignoring the input
Perfect factor recovery	0.704	a model fitted to the latent factors that generated each case
Teacher self-agreement	0.735	a fresh teacher sample against gold built from the others

Scores well above 0.735 mean a model is learning the teacher's quirks. Per-question ceilings vary from 0.560 (agent_trace/urgency) to 0.937 (customer_service/category), so read scores per question as well as on average.

Specialist and general numbers are not comparable. A specialist is fitted on train for these four workflows and cannot answer anything else. A general model takes any question schema at request time and has never seen these. Say which mode you used; the gap between them is the price of generality, not a quality ranking.

The data

Type	Answer	Shape
noul	yes/no	probability that the statement is true
choice	one of N labels	distribution over labels, plus confidence
score	an ordered rubric	distribution over levels, plus an expected score

Every option carries a written description in criteria, and the descriptions are part of the input.

Workflow	Decision	Train	Test
agent_trace_observability	Does an agent run need human review, and how urgently?	300	100
customer_service	The right response and action for a customer thread and account.	300	100
invoice_processing	Pay, hold or reject a vendor bill against its order and delivery.	300	100
security_incidents	Close, investigate or contain a security alert, given machine history.	300	100

state and questions together are exactly the body of a POST /v1/systemone request. gold holds the full gold distributions, and flat <question>__label / __probabilities / __score / __probability_true columns hold the same answers for convenience. factors and label_agreement describe how the case was built and are not model input.

from datasets import load_dataset
import json

ds = load_dataset("LocalLLaMA/typed-decisions", "customer_service", split="test")
row = ds[0]
state, questions, gold = (json.loads(row[k]) for k in ("state", "questions", "gold"))
print(gold["urgency"]["probabilities"])   # score against the full distribution

Report KL or log loss and Brier next to accuracy; calibration is the point.

How it was built

Each case starts from independently sampled latent factors (topic, tone, severity, discrepancy type and so on), which a model renders into free text where the artefact is textual and keeps structured where it is structured. A teacher labels each case three times at temperature 0.7, and the gold is the mean of those distributions, so it stays soft where a decision is genuinely ambiguous. Before release, an audit checks state diversity, label balance, and that the gold actually tracks the input. train comes from a separate run at a different seed; packaging refuses to build if any case id or state appears in both splits.

The write-up behind the benchmark, including the two bugs it exposed in the classifier library: Typed Decisions on Latent Node.