apsua/MIRON_Benchmark

M.I.R.O.N. (Multi-aspect Inference Robustness on Objective Next-tokens)

M.I.R.O.N. is a specialized benchmark designed to evaluate the impact of tokenization and architectural constraints on the generation quality of small, Base language models (SLMs).

Unlike global benchmarks (MMLU, GSM8K), MIRON focuses on the atomic capabilities of a model: morphological generalization, noise robustness, and factual integrity within a simple next-token prediction task.

🎯 Main Goal

To evaluate not the "intelligence" in complex reasoning, but the fundamental capability to correctly process input data, robustness to word fragmentation by the tokenizer, and prediction stability under noise conditions.

The benchmark is designed to be solvable even by small transformers. The tasks do not require Instruction Following, making this dataset ideal for testing Pre-trained (Base) checkpoints.

🧩 Data Structure

The dataset consists of 4000 examples (1000 per category), separated into two languages (ru, en). The dataset uses short category tags:

Tag (Category)	Full Name	Description
Morphology	Morphological Generalization	Tests on pseudo-words (wug-words). Checks if the model can inflect non-existent words (e.g., «wug» -> «wugs») based solely on grammar, without relying on lexical memory.
Facts	Factual Knowledge	Control group. Checks the integrity of perception regarding common named entities (e.g., «Paris», «Sun»). If the tokenizer splits them poorly, access to knowledge becomes difficult.
Logic	Logical Patterns	Simple numeric and algorithmic sequences (e.g., «Tuesday -> Wednesday»). Assesses token stitching when working with numbers and logic.
Noise	Noise Robustness	The context contains typos and perturbations. Evaluates how much the model's confidence "drifts" with slight input distortions.

📊 Dataset Fields

• prefix: Input context for the model.
• target: The expected continuation (ground truth).
• category: Test category (Morphology, Facts, Logic, Noise).

📐 Evaluation Methodology (Metrics)

Two metrics are calculated for each example. This allows distinguishing a model that "does not know" (low Score) from a model that "doubts due to tokenization" (low Confidence).

1. Levenshtein Score (Generation Quality): Normalized Levenshtein distance. Evaluates how close the generated text is to the reference. Range: 0.0% – 100.0%

$${\text{Score}}_c=\frac{1}{\mathrm{\mid }{D}_c\mathrm{\mid }}\sum _{i=1}^{\mathrm{\mid }{D}_c\mathrm{\mid }}(1-\frac{\text{Lev}(S_{\text{gen}}^{(i)},S_{\text{ref}}^{(i)})}{\max (\mathrm{\mid }{S}_{\text{gen}}^{(i)}\mathrm{\mid },\mathrm{\mid }{S}_{\text{ref}}^{(i)}\mathrm{\mid })})\times 100\mathrm{\%}\backslash text\{Score\}\_c\; =\; \backslash frac\{1\}\{|D\_c|\}\; \backslash sum\_\{i=1\}^\{|D\_c|\}\; \backslash left(\; 1\; -\; \backslash frac\{\backslash text\{Lev\}(S\_\{\backslash text\{gen\}\}^\{(i)\},\; S\_\{\backslash text\{ref\}\}^\{(i)\})\}\{\backslash max(|S\_\{\backslash text\{gen\}\}^\{(i)\}|,\; |S\_\{\backslash text\{ref\}\}^\{(i)\}|)\}\; \backslash right)\; \backslash times\; 100\backslash \%$$

2. Target Confidence (Ground Truth Certainty): The geometric mean probability of the tokens that make up the actual target. It shows how ready the model was to output the correct answer. Range: 0.0% – 100.0%

💻 Evaluation Code Example (Python)

import torch
import numpy as np
from Levenshtein import distance as lev_distance

def compute_metrics(model, tokenizer, prefix: str, target: str, generated_text: str, device='cuda'):
    model.eval()
    
    # 1. Levenshtein Score
    max_len = max(len(generated_text), len(target))
    if max_len == 0:
        lev_score = 100.0
    else:
        dist = lev_distance(generated_text, target)
        lev_score = (1 - dist / max_len) * 100.0

    # 2. Target Confidence
    prefix_ids = tokenizer(prefix, return_tensors="pt").input_ids.to(device)
    full_ids = tokenizer(prefix + target, return_tensors="pt").input_ids.to(device)
    
    prefix_len = prefix_ids.shape[1]
    target_len = full_ids.shape[1] - prefix_len
    
    if target_len <= 0:
        return {'lev_score': round(lev_score, 2), 'target_confidence': 0.0}

    with torch.no_grad():
        logits = model(full_ids).logits
    
    shift_logits = logits[0, prefix_len-1:-1, :]
    target_labels = full_ids[0, prefix_len:]
    
    log_probs = torch.log_softmax(shift_logits, dim=-1)
    target_log_probs = torch.gather(log_probs, 1, target_labels.unsqueeze(1)).squeeze()
    
    if target_log_probs.dim() == 0:
        target_log_probs = target_log_probs.unsqueeze(0)

    confidence = np.exp(target_log_probs.mean().item()) * 100.0

    return {
        'lev_score': round(lev_score, 2),
        'target_confidence': round(confidence, 2)
    }