Upload 3 files
Browse files- README.md +111 -3
- miron_en.parquet +3 -0
- miron_ru.parquet +3 -0
README.md
CHANGED
|
@@ -1,3 +1,111 @@
|
|
| 1 |
-
---
|
| 2 |
-
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- ru
|
| 4 |
+
- en
|
| 5 |
+
license: mit
|
| 6 |
+
task_categories:
|
| 7 |
+
- text-generation
|
| 8 |
+
- fill-mask
|
| 9 |
+
tags:
|
| 10 |
+
- tokenization
|
| 11 |
+
- robustness
|
| 12 |
+
- morphology
|
| 13 |
+
- evaluation
|
| 14 |
+
- base-models
|
| 15 |
+
size_categories:
|
| 16 |
+
- 1K<n<10K
|
| 17 |
+
---
|
| 18 |
+
|
| 19 |
+
# M.I.R.O.N. (Multi-aspect Inference Robustness on Objective Next-tokens)
|
| 20 |
+
|
| 21 |
+
**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).
|
| 22 |
+
|
| 23 |
+
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.
|
| 24 |
+
|
| 25 |
+
## 🎯 Main Goal
|
| 26 |
+
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.
|
| 27 |
+
|
| 28 |
+
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.
|
| 29 |
+
|
| 30 |
+
## 🧩 Data Structure
|
| 31 |
+
|
| 32 |
+
The dataset consists of **4000 examples** (1000 per category), separated into two languages (`ru`, `en`). The dataset uses short category tags:
|
| 33 |
+
|
| 34 |
+
| Tag (Category) | Full Name | Description |
|
| 35 |
+
| :--- | :--- | :--- |
|
| 36 |
+
| **`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. |
|
| 37 |
+
| **`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. |
|
| 38 |
+
| **`Logic`** | Logical Patterns | Simple numeric and algorithmic sequences (e.g., *«Tuesday -> Wednesday»*). Assesses token stitching when working with numbers and logic. |
|
| 39 |
+
| **`Noise`** | Noise Robustness | The context contains typos and perturbations. Evaluates how much the model's confidence "drifts" with slight input distortions. |
|
| 40 |
+
|
| 41 |
+
## 📊 Dataset Fields
|
| 42 |
+
* **`prefix`**: Input context for the model.
|
| 43 |
+
* **`target`**: The expected continuation (ground truth).
|
| 44 |
+
* **`category`**: Test category (`Morphology`, `Facts`, `Logic`, `Noise`).
|
| 45 |
+
|
| 46 |
+
## 📐 Evaluation Methodology (Metrics)
|
| 47 |
+
|
| 48 |
+
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).
|
| 49 |
+
|
| 50 |
+
1. **Levenshtein Score (Generation Quality):**
|
| 51 |
+
Normalized Levenshtein distance. Evaluates how close the generated text is to the reference.
|
| 52 |
+
*Range: 0.0% – 100.0%*
|
| 53 |
+
|
| 54 |
+
$$
|
| 55 |
+
\text{Score}_c = \frac{1}{|D_c|} \sum_{i=1}^{|D_c|} \left( 1 - \frac{\text{Lev}(S_{\text{gen}}^{(i)}, S_{\text{ref}}^{(i)})}{\max(|S_{\text{gen}}^{(i)}|, |S_{\text{ref}}^{(i)}|)} \right) \times 100\%
|
| 56 |
+
$$
|
| 57 |
+
|
| 58 |
+
2. **Target Confidence (Ground Truth Certainty):**
|
| 59 |
+
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.
|
| 60 |
+
*Range: 0.0% – 100.0%*
|
| 61 |
+
|
| 62 |
+
$$
|
| 63 |
+
\text{Conf}(S_{\text{ref}}) = \exp \left( \frac{1}{T} \sum_{j=1}^{T} \log P(t_j \mid \text{context}, t_{<j}) \right) \times 100\%
|
| 64 |
+
$$
|
| 65 |
+
|
| 66 |
+
### 💻 Evaluation Code Example (Python)
|
| 67 |
+
|
| 68 |
+
```python
|
| 69 |
+
import torch
|
| 70 |
+
import numpy as np
|
| 71 |
+
from Levenshtein import distance as lev_distance
|
| 72 |
+
|
| 73 |
+
def compute_metrics(model, tokenizer, prefix: str, target: str, generated_text: str, device='cuda'):
|
| 74 |
+
model.eval()
|
| 75 |
+
|
| 76 |
+
# 1. Levenshtein Score
|
| 77 |
+
max_len = max(len(generated_text), len(target))
|
| 78 |
+
if max_len == 0:
|
| 79 |
+
lev_score = 100.0
|
| 80 |
+
else:
|
| 81 |
+
dist = lev_distance(generated_text, target)
|
| 82 |
+
lev_score = (1 - dist / max_len) * 100.0
|
| 83 |
+
|
| 84 |
+
# 2. Target Confidence
|
| 85 |
+
prefix_ids = tokenizer(prefix, return_tensors="pt").input_ids.to(device)
|
| 86 |
+
full_ids = tokenizer(prefix + target, return_tensors="pt").input_ids.to(device)
|
| 87 |
+
|
| 88 |
+
prefix_len = prefix_ids.shape[1]
|
| 89 |
+
target_len = full_ids.shape[1] - prefix_len
|
| 90 |
+
|
| 91 |
+
if target_len <= 0:
|
| 92 |
+
return {'lev_score': round(lev_score, 2), 'target_confidence': 0.0}
|
| 93 |
+
|
| 94 |
+
with torch.no_grad():
|
| 95 |
+
logits = model(full_ids).logits
|
| 96 |
+
|
| 97 |
+
shift_logits = logits[0, prefix_len-1:-1, :]
|
| 98 |
+
target_labels = full_ids[0, prefix_len:]
|
| 99 |
+
|
| 100 |
+
log_probs = torch.log_softmax(shift_logits, dim=-1)
|
| 101 |
+
target_log_probs = torch.gather(log_probs, 1, target_labels.unsqueeze(1)).squeeze()
|
| 102 |
+
|
| 103 |
+
if target_log_probs.dim() == 0:
|
| 104 |
+
target_log_probs = target_log_probs.unsqueeze(0)
|
| 105 |
+
|
| 106 |
+
confidence = np.exp(target_log_probs.mean().item()) * 100.0
|
| 107 |
+
|
| 108 |
+
return {
|
| 109 |
+
'lev_score': round(lev_score, 2),
|
| 110 |
+
'target_confidence': round(confidence, 2)
|
| 111 |
+
}
|
miron_en.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c9d95c3f82a5dc8ac3e3481dca88d904bc5678cd4545ea2d5e371314d78cd2ab
|
| 3 |
+
size 97713
|
miron_ru.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5df2078aa0c30efe1a09cd4c460c69661b5d66e8f3a859ffb92e1b828cbc1997
|
| 3 |
+
size 153034
|