apsua/MIRON_Benchmark

M.I.R.O.N. Benchmark

M.I.R.O.N. (Multi-aspect Inference Robustness on Objective Next-tokens) is a Russian and English benchmark for examining how language models continue short prompts. It focuses on morphology, factual knowledge, simple patterns and robustness to noisy input. The tasks are designed for base models and do not require instruction following.

Dataset

The benchmark contains 4,000 examples per language, with 1,000 examples per category in each language. Select ru or en in the Dataset Viewer to inspect a language.

Category	What it probes
Morphology	Inflection and generalization using pseudowords.
Facts	Continuations involving familiar facts and named entities.
Logic	Simple numerical and sequential patterns.
Noise	Continuation when the input contains perturbations or typos.

Each row has three fields:

Field	Description
prefix	Input text provided to the model.
target	Reference continuation.
category	One of the four categories above.

Loading

from datasets import load_dataset

russian = load_dataset("apsua/MIRON_Benchmark", "ru", split="train")
english = load_dataset("apsua/MIRON_Benchmark", "en", split="train")

Evaluation

The benchmark was designed around two complementary measures:

• Generation score: character-level Levenshtein similarity between a generated continuation and the reference, expressed as a percentage. For two empty strings, the similarity is 100%.
• Target confidence: geometric mean of the model probabilities assigned to the reference continuation tokens, conditioned on the prefix and preceding reference tokens, expressed as a percentage.

License

MIT, as specified in the dataset metadata.