HumanNumberEval

HumanNumberEval is a compact benchmark for testing whether AI systems can distinguish human-chosen numbers from numbers produced by a random number generator.

The benchmark contains 200 labeled examples:

• 100 human numbers
• 100 rng numbers

Each example is a single number with a binary label. The model sees one number at a time and must answer with exactly human or rng.

Dataset Files

This upload contains only the files needed to use and inspect the benchmark:

• numbers.csv
• results.csv: model prediction/evaluation records
• model_scores.png: model score visualization
• README.md: dataset card
• LICENSE: Apache License 2.0

Dataset Schema

numbers.csv has two columns:

Column	Type	Description
number	string/integer	The number to classify. Treat as a string if preserving formatting matters.
label	string	Either human or rng.

Example rows:

number,label
37,human
73,human
832,human
181,rng
585,rng

Task

Given a number, predict one of:

• human: the number was chosen by a human
• rng: the number was produced by a random number generator

This is a binary classification benchmark. Random guessing is expected to score around 50% accuracy because the dataset is balanced.

Submit Your Own Numbers

If you want to submit your own numbers for future versions of this benchmark, use this form:

https://docs.google.com/forms/d/e/1FAIpQLSclPrTaO5c59JIuOkb8XGygF2R2bbmdPhy5-6hpDF4NJ7b0zw/viewform

Prompt

The benchmark prompt used for model evaluation was:

Is this number human or RNG The number is {number} you need to just reply with "human" or "rng" nothing else

{number} is replaced with the value from the number column.

Loading With datasets

from datasets import load_dataset

dataset = load_dataset(
    "csv",
    data_files={"test": "numbers.csv"},
)

print(dataset["test"][0])

If you upload this repository as a Hugging Face dataset, it can be loaded by replacing the local path with the dataset repository name:

from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/HumanNumberEval")

print(dataset["test"][0])

To load the model evaluation records instead:

from datasets import load_dataset

results = load_dataset("YOUR_USERNAME/HumanNumberEval", "results")

print(results["test"][0])

Evaluation

Recommended metric:

• Accuracy

Because the label distribution is balanced, random guessing should score around 50% accuracy. The included results.csv file contains reference model outputs evaluated with the prompt above.

Model Results

The error columns show directional mistakes: RNG predicted human counts random numbers labeled as human, and Human predicted RNG counts human numbers labeled as random. Predicted human/RNG shows each model's valid-label prediction split, while Other/error counts invalid outputs or failed calls.

A number of models are close to 50% accuracy because they show a strong label bias rather than balanced discrimination. For example, some models predict almost everything as human, while others predict almost everything as rng.

Rank	Model	Accuracy	Correct	RNG predicted human	Human predicted RNG	Predicted human/RNG	Other/error
1	z-ai/glm-5.1	67.0%	134/200	37	25	110/86	4
2	moonshotai/kimi-k2.6	64.5%	129/200	24	47	77/123	0
3	z-ai/glm-5.2	57.5%	115/200	44	38	105/92	3
4	google/gemini-2.5-pro	55.5%	111/200	59	30	129/71	0
5	openai/gpt-oss-120b	51.5%	103/200	80	16	164/35	1
6	openai/gpt-5.5	51.0%	102/200	56	42	114/86	0
7	anthropic/claude-opus-4.8	50.5%	101/200	4	93	10/188	2
8	deepseek/deepseek-v4-pro	50.5%	101/200	92	4	187/10	3
9	google/gemma-4-26b-a4b-it	50.5%	101/200	11	88	23/177	0
10	google/gemma-4-31b-it	50.5%	101/200	1	98	3/197	0
11	mistralai/mistral-medium-3-5	50.5%	101/200	0	99	1/199	0
12	anthropic/claude-haiku-4.5	50.0%	100/200	100	0	200/0	0
13	anthropic/claude-opus-4.6	50.0%	100/200	100	0	200/0	0
14	anthropic/claude-sonnet-4.6	50.0%	100/200	100	0	200/0	0
15	google/gemini-3-flash-preview	50.0%	100/200	100	0	200/0	0
16	meta-llama/llama-3.3-70b-instruct	50.0%	100/200	100	0	200/0	0
17	x-ai/grok-4.3	50.0%	100/200	0	100	0/200	0
18	deepseek/deepseek-v4-flash	49.0%	98/200	93	8	184/15	1
19	z-ai/glm-5	49.0%	98/200	86	16	170/30	0
20	google/gemini-2.5-flash-lite	48.5%	97/200	8	95	13/187	0
21	google/gemini-3.5-flash	47.0%	94/200	98	8	190/10	0
22	google/gemini-3.1-flash-lite	46.5%	93/200	9	98	11/189	0
23	minimax/minimax-m3	46.0%	92/200	33	72	59/138	3
24	openai/gpt-3.5-turbo	46.0%	92/200	98	10	188/12	0
25	google/gemini-2.5-flash	45.5%	91/200	18	91	27/173	0
26	qwen/qwen3.7-plus	44.5%	89/200	41	70	71/129	0
27	google/gemini-3.1-pro-preview	42.5%	85/200	67	43	121/74	5
28	qwen/qwen3.6-27b	38.0%	76/200	91	33	158/42	0
29	anthropic/claude-opus-4.7	37.0%	74/200	92	34	158/42	0
30	qwen/qwen3.6-plus	36.0%	72/200	71	57	114/86	0
31	openai/gpt-4o	34.0%	68/200	51	81	70/130	0
32	openai/gpt-4.1	23.0%	46/200	99	55	144/56	0

License

This dataset is released under the Apache License 2.0. See LICENSE for the full license text.

Citation

If you use this benchmark, cite the dataset repository:

@dataset{humannumbereval,
  title = {HumanNumberEval},
  description = {A benchmark for classifying human-chosen numbers versus RNG-generated numbers},
  year = {2026}
}