LvW-bench

A benchmark for tiny models that you run on your own machine. It scores both kinds of small model in one place:

• System 1, decide: typed decisions. The model gets a state and typed questions (choice, score, noul) and answers each with a probability distribution. Scores are accuracy, KL from gold, Brier and ECE, the same as LocalLLaMA/typed-decisions. Native decision models (d1, Jev-style heads) answer through /v1/systemone. A text model answers through the probabilities of its first output token.
• System 2, answer and reason: generated answers on permissive subsets of GSM8K, MMLU, MMLU-Pro, TruthfulQA and IFEval, plus a seeded generated set. Each model runs twice, once with thinking off and a small budget (answer) and once with thinking on (reason). Thinking is native <think> when the chat template has it, and step by step otherwise.

The runner is lvwb, from the GitHub repo. It runs on macOS, Linux and Windows, from the terminal (CLI and TUI), from a desktop app, or in the browser. It accepts a Hugging Face link or id, a local .gguf file, a local .safetensors folder, or any OpenAI-compatible server.

Leaderboard

Bench version 0.1.0, standard profile:

• decide uses the whole decide_all test split (500 cases, 1,500 decisions).
• answer uses the whole generative_all test split (1,143 items).
• reason uses reason_subset (240 items, 40 per set, 2,048-token budget).

All runs were local, on llama.cpp for GGUF and transformers for safetensors. Every row records the exact repo commit, file, quant and engine build. The result and prediction files are in results/.

System 1 · decide

#	Model	Accuracy ↑	KL ↓	Brier ↓	ECE ↓	Coverage	p50 / case	Adapter
1	LiquidAI/d1-2B-GGUF · GGUF Q4_K_M	0.580	0.635	0.360	0.107	100%	1698 ms	systemone
2	Lucasgitguy-co/lfm3-vl-2b-heretic-gguf · GGUF Q4_K_M	0.432	1.914	0.684	0.359	100%	3294 ms	logprobs
3	LiquidAI/LFM2-350M-GGUF · GGUF Q4_K_M	0.410	1.320	0.555	0.240	100%	918 ms	logprobs
4	Lucasgitguy-co/minicpm-mermaid-gguf · GGUF Q4_K_M	0.396	1.606	0.685	0.378	100%	3424 ms	logprobs
5	bartowski/SmolLM2-1.7B-Instruct-GGUF · GGUF Q4_K_M	0.383	0.813	0.479	0.144	100%	3183 ms	logprobs
6	unsloth/Qwen3-0.6B-GGUF · GGUF Q4_K_M	0.373	1.680	0.752	0.410	100%	1494 ms	logprobs
7	Lucasgitguy-co/minimind-mermaid-run2/ep3 · safetensors	0.353	2.669	0.631	0.347	79%	335 ms	logprobs
8	LiquidAI/LFM2-1.2B-GGUF · GGUF Q4_K_M	0.299	1.980	0.798	0.489	100%	2286 ms	logprobs
9	bartowski/SmolLM2-135M-Instruct-GGUF · GGUF Q4_K_M	0.282	0.888	0.519	0.224	100%	175 ms	logprobs
10	bartowski/SmolLM2-360M-Instruct-GGUF · GGUF Q4_K_M	0.271	0.909	0.505	0.203	100%	577 ms	logprobs
–	Prior (ignores the input)	0.482	0.631	0.373	0.022	–	–	reference
–	Uniform (same probability on every option)	0.331	0.709	0.429	0.105	–	–	reference

System 1 · decide: specialists

Trained on the same task as a workflow (the Spot Snake models learned spot_snake moves with the same teacher), so not comparable with the zero-shot table.

#	Model	Accuracy ↑	KL ↓	Brier ↓	ECE ↓	Coverage	p50 / case	Adapter
1	Lucasgitguy-co/spot-snake-d1-3B-v0.1 · GGUF Q4_K_M	0.563	2.372	0.487	0.272	100%	2172 ms	systemone
2	Lucasgitguy-co/spot-snake-d1-2B-v0.1 · GGUF Q4_K_M	0.535	1.722	0.565	0.329	100%	1564 ms	systemone
3	Lucasgitguy-co/spot-snake-minimind-jev-v0.1 · safetensors	0.334	0.709	0.429	0.102	7%	0 ms	systemone

System 2 · answer

#	Model	Accuracy ↑	95% CI	gsm8k	mmlu	mmlu_pro	truthfulqa	ifeval	generated	tok/answer	tok/s
1	Lucasgitguy-co/minicpm-mermaid-gguf · GGUF Q4_K_M	0.396	0.368–0.425	0.080	0.535	0.186	0.360	0.710	0.413	27	57
2	LiquidAI/LFM2-1.2B-GGUF · GGUF Q4_K_M	0.385	0.355–0.413	0.010	0.504	0.150	0.333	0.755	0.449	74	159
3	Lucasgitguy-co/lfm3-vl-2b-heretic-gguf · GGUF Q4_K_M	0.382	0.355–0.410	0.045	0.408	0.186	0.360	0.735	0.480	54	91
4	bartowski/SmolLM2-1.7B-Instruct-GGUF · GGUF Q4_K_M	0.305	0.279–0.333	0.030	0.548	0.171	0.213	0.530	0.249	64	82
5	unsloth/Qwen3-0.6B-GGUF · GGUF Q4_K_M	0.298	0.273–0.326	0.025	0.386	0.136	0.320	0.605	0.267	57	129
6	LiquidAI/LFM2-350M-GGUF · GGUF Q4_K_M	0.269	0.241–0.293	0.020	0.447	0.121	0.220	0.565	0.169	89	346
7	bartowski/SmolLM2-360M-Instruct-GGUF · GGUF Q4_K_M	0.212	0.187–0.236	0.045	0.259	0.136	0.207	0.390	0.204	82	233
8	bartowski/SmolLM2-135M-Instruct-GGUF · GGUF Q4_K_M	0.158	0.137–0.181	0.015	0.228	0.100	0.213	0.225	0.156	110	749
9	Lucasgitguy-co/minimind-mermaid-run2/ep3 · safetensors	0.029	0.019–0.039	0.010	0.009	0.007	0.000	0.105	0.031	193	399

System 2 · reason

#	Model	Accuracy ↑	95% CI	gsm8k	mmlu	mmlu_pro	truthfulqa	ifeval	generated	tok/answer	tok/s
1	LiquidAI/LFM2-1.2B-GGUF · GGUF Q4_K_M	0.517	0.450–0.579	0.600	0.500	0.300	0.375	0.750	0.575	461	206
2	Lucasgitguy-co/lfm3-vl-2b-heretic-gguf · GGUF Q4_K_M	0.508	0.450–0.571	0.700	0.500	0.250	0.375	0.450	0.775	1477	140
3	Lucasgitguy-co/minicpm-mermaid-gguf · GGUF Q4_K_M	0.404	0.338–0.463	0.100	0.475	0.350	0.425	0.650	0.425	488	168
4	bartowski/SmolLM2-1.7B-Instruct-GGUF · GGUF Q4_K_M	0.379	0.317–0.442	0.400	0.500	0.275	0.200	0.425	0.475	316	103
5	LiquidAI/LFM2-350M-GGUF · GGUF Q4_K_M	0.317	0.258–0.375	0.425	0.400	0.200	0.125	0.500	0.250	350	502
6	unsloth/Qwen3-0.6B-GGUF · GGUF Q4_K_M	0.312	0.254–0.371	0.650	0.075	0.025	0.025	0.650	0.450	924	187
7	bartowski/SmolLM2-135M-Instruct-GGUF · GGUF Q4_K_M	0.146	0.104–0.192	0.025	0.150	0.225	0.150	0.225	0.100	954	603
8	bartowski/SmolLM2-360M-Instruct-GGUF · GGUF Q4_K_M	0.129	0.087–0.171	0.075	0.075	0.100	0.050	0.250	0.225	507	203
9	Lucasgitguy-co/minimind-mermaid-run2/ep3 · safetensors	0.033	0.013–0.058	0.000	0.000	0.000	0.000	0.175	0.025	2202	370

To add a model, run it with lvwb run <model> and open a discussion. Attach the result JSON and its .predictions.jsonl, and say whether the model is general or a specialist. lvwb submit <result.json> prints the steps. bench/score.py rescores the predictions file to the same numbers without the model.

Benchmarking a model

brew trust gitguy268/tap && brew install gitguy268/tap/lvw-bench   # or a release binary
lvwb run Lucasgitguy-co/spot-snake-d1-2B-v0.1           # HF id: picks the GGUF, starts llama-server
lvwb run https://huggingface.co/unsloth/Qwen3-0.6B-GGUF # HF link
lvwb run ~/models/my-model-Q4_K_M.gguf                  # local GGUF
lvwb run ~/models/my-model/                             # local safetensors folder (uv + transformers)
lvwb run http://127.0.0.1:8080/v1 --backend openai      # a server that is already running
lvwb tui                                                # pick a preset in the terminal UI

• --track decide|answer|reason
• --profile quick|standard|full
• --quant Q8_0, --subfolder ep3, --revision refs/pr/1
• --seed N draws a fresh generated set
• --mode specialist if the model was trained on train

GGUF needs llama-server (brew install llama.cpp). Native /v1/systemone decisions need llama.cpp 0.9.0 or later. Safetensors need uv. The first run installs PyTorch in a cached environment.

Rescoring with Python only

pip install datasets
python bench/score.py my-model__standard__20261105-120000.predictions.jsonl

bench/score.py holds the same rules as the Go scorer, and a golden test file keeps the two identical. The System 1 part follows typed-decisions' bench/score_with_jobs.py.

Reading a score

• System 1 gold is computed by rules, not by a teacher model. Each workflow builds a case from sampled facts and derives the gold distribution from fixed rules. Facts that can be read off the state, such as whether two meetings overlap, have a sharp gold. Judgement calls, such as which action fits a damaged parcel, keep a soft gold. A score measures agreement with those rules. The Prior row, label frequencies from train that ignore the input, is the floor to beat.
• System 2 scores generated answers. The scorer extracts the final answer (the Answer: line, the leading letter or the last number) after removing <think> blocks. A think block that never closes means the budget ran out, and that answer counts as wrong. These numbers are not comparable with lm-eval loglikelihood scores (for example TPN's ARC or MMLU acc_norm), which compare option probabilities and never let the model write.
• General and specialist rows are not comparable. The Spot Snake models were trained on Spot's moves with the same teacher and text as spot_snake, so they are listed as specialists. spot_snake has more "pocket" positions than real games, where one move closes Spot in. The Spot Snake d1 models mostly keep going straight there, so they score lower than their in-game teacher agreement.
• Coverage is the share of questions a model answered. A head that only ranks moves answers the other questions with nothing, and those are scored as uniform.

The data

Config	Track	Train	Test	Source	Licence
spot_snake	decide	300	100	original, rule gold	MIT
it_support	decide	300	100	original, rule gold	MIT
delivery	decide	300	100	original, rule gold	MIT
code_review	decide	300	100	original, rule gold	MIT
calendar	decide	300	100	original, rule gold	MIT
decide_all	decide	1500	500	all five workflows	MIT
gsm8k	answer, reason	–	200	openai/gsm8k test	MIT
mmlu	answer, reason	–	228	cais/mmlu test, 4 per subject	MIT
mmlu_pro	answer, reason	–	140	TIGER-Lab/MMLU-Pro test, 10 per category	MIT
truthfulqa	answer, reason	–	150	truthfulqa/truthful_qa mc1, options shuffled	Apache-2.0
ifeval	answer, reason	–	200	google/IFEval, verifiable subset	Apache-2.0
generated	answer, reason	–	225	original, seeded generator (seed 7)	MIT
generative_all	answer	–	1143	the six sets above	mixed, see rows
reason_subset	reason	–	240	40 per set, the items the reason leaderboard runs	mixed, see rows

System 1 rows

state and questions together are the body of a POST /v1/systemone request. gold holds the gold distribution per question. The flat columns <question>__label, __probabilities, __confidence, __score and __probability_true hold the same values. factors describe how the case was built and are not model input.

Workflow	Questions
spot_snake	move (choice), eats_now (noul), boxed_in_risk (score 0–3)
it_support	category (choice), priority (score), needs_onsite (noul)
delivery	action (choice), notify_customer (noul), severity (score)
code_review	decision (choice), risk (score), needs_tests (noul)
calendar	clash (noul), resolution (choice), urgency (score)

from datasets import load_dataset
import json
row = load_dataset("Lucasgitguy-co/LvW-bench", "decide_all", split="test")[0]
state, questions, gold = (json.loads(row[k]) for k in ("state", "questions", "gold"))

System 2 rows

Each row has prompt, choices (multiple choice only), answer and answer_type. answer_type is one of mc, number, exact, ifeval, json or mermaid. The checker for each type is in bench/score.py. meta holds what the checker needs: IFEval instruction ids and kwargs, required JSON values, or required Mermaid edges.

How it was built

• Sources: permissive subsets are sampled with a fixed seed. ARC (CC-BY-SA) and HellaSwag (no licence) are left out. TruthfulQA options are shuffled, because upstream always lists the correct one first. IFEval keeps only prompts whose instructions can be verified (17 instruction types).
• System 1: train comes from different seeds than test, and the build refuses to package if a state appears in both.
• Generated set: 9 families of 25 items each: arithmetic, word problems, letter counts, sorting, weekdays and dates, syllogisms, list lookups, JSON output and Mermaid flowcharts. Every answer is computed, so --seed makes a fresh set no model can have memorised.

Limits

• The models are small and the sets are small. Read the 95% confidence intervals before ranking two close rows.
• System 1 gold follows simple rules written for this benchmark. A model that reasons better than the rules can score lower.
• Running a text model on System 1 through first-token probabilities depends on the chat template. Backends without logprobs fall back to sampling 8 answers per question, marked sampled.
• Text only in v0.1. Vision-language models run in text mode.

Licence

The original data, code and results are MIT. The rows from GSM8K, MMLU and MMLU-Pro stay MIT, and the rows from TruthfulQA and IFEval stay Apache-2.0. Each row's license column says which applies. See NOTICE and licenses/. Models keep their own licences.

@misc{lvwbench2026,
  title  = {LvW-bench: a local benchmark for tiny System 1 and System 2 models},
  author = {Van De Weerd, Lucas},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Lucasgitguy-co/LvW-bench}
}