SupraLabs/Supra-microBench

Supra-μBench • 512 max seq len • Every small model welcome

Supra-μBench (Supra-microBench) v0.2.0 is a log-likelihood benchmark for very small base language models, built by SupraLabs. It's built for evaluation your small models under 100M parameters (yea, we evaluated Qwen3-0.6B too but that was only for comparison!). A model is ranked by how much probability it assigns to the correct continuation (loglikelihood like you know it from the LM-Eval tool).

The set has about 1060 English items across 18 tasks and six domains. Items are produced by seeded generators plus a smaller set of hand-written items. The generation seed is 1234. Each task has its own random stream, so editing one task does not reshuffle the others.

IMPORTANT

1. Do NOT train on this dataset! It's for evaluation only!!
2. This is for BASE models only - not for instruction tuned models. Maybe we'll release a benchmark for instruct models too...
3. Your model SHOULD have at least 512 tokens of context window

Domains & Tasks

The μBench Intelligence Index is a weighted mean of the six domain scores. Each domain score is the unweighted mean of its task scores. Every score is on a 0 to 100 scale.

Domain	Name	Weight
D1	Linguistic Competence	0.15
D2	Lexicon & Semantics	0.15
D3	World Knowledge	0.15
D4	Context Use & Reading	0.20
D5	Elementary Reasoning	0.20
D6	Coherence, Common Sense & LM	0.15

How a choice item is scored

The evaluator loads a causal language model and its tokenizer. Every sequence starts with a BOS token, or EOS if the tokenizer has no BOS. Continuations are scored by summed log-probability. No tokens are sampled.

Three reductions are used. sum compares raw log-probabilities. mean_byte divides by the UTF-8 length of the continuation so a longer string is not punished. pmi subtracts the log-probability of the same continuation under a null context, so a word that is likely in general does not win just for being frequent.

The predicted choice is the one with the highest score. Accuracy is then chance-normalized:

score = 100 * max(0, (accuracy - chance) / (1 - chance))

A model at random chance scores 0. A model that is always correct scores 100. Scores are not allowed to go below 0 (and what bad model would that be 😭🤣).

On tasks marked with group accuracy, a contrast pair is one unit. Both members have to be correct. That cancels a model that always prefers one surface form. Units that have no group are scored alone.

The bits-per-byte task is separate. Fresh text is scored with a sliding window. Bits per byte are mapped onto 0 to 100 with provisional anchors: a floor of 2.5 (a weak unigram-like model) and a ceiling of 0.5 (a strong big LLM like a 2B or 4B model). The mapping is logarithmic and clipped to 0 to 100.

Reported standard errors are item-sampling errors. Treat an index gap as noise when it is smaller than 2 times the square root of the sum of the two squared standard errors.

How to submit a run

Run the SupraLabs evaluator on a local model folder or a Hub model id. The folder must contain the tokenizer.

That JSON is the leaderboard source. It stores the model path, the parameter count, per-task scores, the six domain scores, and mubench_intelligence_index.

You can find evaluate.py and tasks.py in the files list of the repo.

Example of how to run the evaluation:

python3 evaluate.py --model_dir SupraLabs/SupraNeo-4M --trust_remote_code --show_levels --output SupraNeo4M-Eval-Results.json

Leaderboard

One run is on file. Rank is by INDEX, high to low.

Rank	Modelname	Org	Params	INDEX	D1	D2	D3	D4	D5	D6
🥇 1	Qwen3-0.6B	Qwen	596.0M	69.82	59.82	79.42	87.94	64.03	77.10	50.10
🥈 2	SmolLM2-135M-Base	HuggingFaceTB	134.5M	58.90	62.18	86.33	86.85	40.85	45.93	41.62
🥉 3	Supra2-100M-Base	SupraLabs	100.7M	45.38	58.14	73.42	64.67	26.62	21.75	41.77
4	Pythia-160M	EleutherAI	162.3M	42.47	58.69	72.57	48.51	25.48	26.79	33.69
5	Supra-50M-Base	SupraLabs	51.8M	40.11	45.11	75.34	60.57	30.20	12.54	29.40
6	Supra2-Medium-Base	SupraLabs	25.4M	26.70	0.44	57.07	49.50	19.50	14.35	25.84
7	Pythia-70M	EleutherAI	70.4M	25.45	24.57	48.26	34.54	14.69	18.18	18.47
8	ForgePlex-M2-9M	ForgeWorks	9.9M	23.78	31.47	34.27	22.97	22.38	15.08	19.85
9	BananaMind-2-Nano	BananaMind	10.0M	21.51	11.25	46.03	27.90	17.31	14.96	15.21
10	NCN-RW	AxiomicLabs	9.7M	18.99	16.26	33.53	18.52	18.13	12.63	17.27
11	Spark-2A	SurjoLabs	9.1M	18.86	13.96	42.59	15.97	16.46	12.80	14.20
12	Pythia-14M	EleutherAI	14.1M	15.10	4.00	33.19	23.44	4.84	13.80	15.19
13	StorySupra-10M	SupraLabs	12.6M	14.46	4.36	13.90	0.80	34.15	5.15	24.97
14	pulvis-v2	bench-labs	3.0M	13.94	3.11	33.19	6.10	12.22	16.27	12.52
15	BananaMind-2-Micro	BananaMind	2.9M	12.04	7.35	16.58	3.81	15.89	14.29	12.25
16	Supra-Mini-v5-8M	SupraLabs	7.9M	11.95	0.00	28.17	19.50	9.22	5.86	11.85
17	Tokle-3M	techdotus	2.9M	11.89	0.00	24.05	9.36	10.32	10.50	18.07
18	Supra-Mini-v3-0.5M	SupraLabs	467648	3.05	0.00	3.70	0.72	1.55	5.88	5.98

Note: we did NOT rank every model that is ranked on the Open SLM Leaderboard. If you have new ranks, you can post them in your model card.

Intended use

Use the index to compare small base models on grammar, lexicon, facts, reading, elementary reasoning, and local coherence. Do not use it as a chat, instruction-following, or safety score. DO NOT TRAIN ON THE TASKS!!!

Limitations

Language: English only. Several tasks have only a few units, so standard errors are sometimes wide. World-knowledge items reflect widely taught facts, not a neutral sample of the world. The bits-per-byte anchors are provisional. A model whose context window is shorter than an item is left-truncated, which can hurt the reading tasks but we tried to keep the tasks inside a 512 token window.

Credits

Supra-μBench was built by SupraLabs. Version 0.2.0.