LLMango responses

This is the dataset behind LLMango: a blog dedicated to showcasing different behaviours of Large Language Models through prompting them in great volumes and data visualization. LLMango is divided into experiments, where each one is different and tries to answer specific questions you might have about LLMs.

One row is one call. It holds the prompt that produced it, the provider's verbatim response, and the usage, cost and timing the provider reported.

• Charts and write-ups: https://llmango.rafalkwiecien.com
• Pipeline that produced this: https://github.com/kwiecien-rafal/llmango

Usage

A config is an experiment. A split is one question inside it.

from datasets import load_dataset

baseline = load_dataset("rafalkwiecien/llmango", "e001_fruit", split="001a")

Every question shares its experiment's columns, so splits within a config compare directly. A later experiment adds the columns its own analysis needs.

Experiments

001: fruit

Experiment: Can an LLM's response meaningfully change when you prompt it in a different language, or when you present information to the model in a different manner?

gpt-5.6-luna at temperature 1.0 is prompted to pick one fruit out of ten, at random. An arm is one combination of prompt language and a response format. There is seventeen arms in total, each sampled 2 000 times, for 34 000 rows. Each split is one question.

Split	Languages	Response schema	List order
001a	en, pl, ja	English schema	fixed
001b	en, pl, ja	English schema	a second fixed order
001c	en, pl, ja	English schema	shuffled per sample
001d	en, pl, ja	English schema, native schema, no schema	shuffled per sample

001a is the baseline distribution. 001b re-asks it in all three languages under a different fixed order, which separates a preference for a fruit from a preference for a position. 001c reshuffles the list for every sample, which removes position as a confound within a language. 001d asks each language under the English schema, under that language's own schema, and under no schema at all, leaving the response schema as the only variable. English contributes two arms rather than three, because its native schema is the English schema. Every 001d prompt adds a "one fruit name only" instruction, because the no-schema arm has nothing else constraining it.

A native schema's name is ASCII, because the provider constrains it: the Japanese one is written in romaji, and the Polish one drops its diacritic.

A shuffle draws from a per-sample seed picked at random by the run. Every arm of a question shares that seed for a given sample, so the order shown is controlled rather than varying alongside the language. No seed is replayed across runs; the order each row actually saw is recorded in prompt_inputs.

Columns

Identity

Column	Type	Meaning
question_id	string	001a … 001d; the only identifier the pipeline takes
lang	string	Language of the prompt
model	string	Model as requested
provider	string	Provider backend that made the call
run_id	string	<question_id>__<UTC timestamp>, one per llmango run
sample_idx	int64	Sample number within the run; shared across arms
temperature	float64	Sampling temperature the question declares

Prompt

Column	Type	Meaning
prompt_sha256	string	Hash of the exact prompt sent
prompt	string	The exact prompt sent
prompt_inputs	string	JSON of the resolved inputs, including the fruit order shown

Answer

Column	Type	Meaning
raw_json	string	Verbatim model output; free text on the no-schema arm
answer	string	The answer field read off the parsed response
canonical	string	Category the answer normalizes to; null when invalid
is_valid	boolean	Whether the answer named something on the list
chosen_position	int64	1-based place of the pick in the list that row saw

Provider response

Column	Type	Meaning
model_snapshot	string	Dated snapshot the provider actually served
finish_reason	string	Why generation stopped
refusal	string	Provider refusal, null when there was none
error	string	Error text, null when the call succeeded
response_id	string	Provider-side id of the completion
service_tier	string	Service tier the call was served on
provider_created_at	timestamp[us, UTC]	Creation time the provider reported
response_schema	string	JSON Schema sent for this arm; null for no-schema
request_envelope	string	Request body
response_envelope	string	Verbatim response body

Usage and cost

Column	Type	Meaning
prompt_tokens	int64	Input tokens
completion_tokens	int64	Output tokens
total_tokens	int64	Sum the provider reported
cached_tokens	int64	Input tokens served from cache
reasoning_tokens	int64	Reasoning tokens, where the model reports them
input_cost_usd	float64	Input cost at the pricing version below
output_cost_usd	float64	Output cost at the pricing version below
total_cost_usd	float64	Sum of the two
pricing_version	string	Date of the price table used

Timing

Column	Type	Meaning
generation_seconds	float64	Wall-clock duration of the call
created_at	timestamp[us, UTC]	When the call was made

Provenance and limitations

• canonical on the no-schema arm is assigned by normalize, which matches offline first and falls back to an LLM for what it cannot resolve.
• A code run appends to a split, without any replacement of already existing data, nothing is deduplicated.
• No personal data and no human subjects, every row is a model's output to a prompt in prompts/.

Licence

CC BY 4.0. Attribution to Rafał Kwiecień, linking back to this dataset.

The model outputs themselves are the raw material; the collection, prompts, normalization and structure are the contribution being licensed.

Citation

@misc{kwiecien_llmango,
  author = {Rafał Kwiecień},
  title  = {LLMango: visualizing AI behaviour},
  url    = {https://huggingface.co/datasets/rafalkwiecien/llmango},
  note   = {Charts and write-ups at https://llmango.rafalkwiecien.com}
}