Update README.md
Browse files
README.md
CHANGED
|
@@ -103,3 +103,47 @@ outputs = model.generate(
|
|
| 103 |
do_sample=False,
|
| 104 |
)
|
| 105 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 103 |
do_sample=False,
|
| 104 |
)
|
| 105 |
```
|
| 106 |
+
# Evaluation
|
| 107 |
+
## Per-Character Perplexity
|
| 108 |
+
**What is Perplexity?** Perplexity measures how well a language model predicts text. A model with low perplexity makes accurate predictions consistently, while a high perplexity means the model is frequently "surprised" by unexpected words or patterns. Lower perplexity indicates the model has learned language patterns more effectively. It's less "surprised" by what it encounters because it better understands how the language works.
|
| 109 |
+
**Why Character-Level?** Different language models use different internal vocabularies - some break text into whole words, others into word fragments, and some into individual characters. This makes direct comparison difficult.
|
| 110 |
+
Character-level perplexity creates a standardised comparison by calculating how well each model would theoretically perform if we measured their predictions character-by-character. We're not changing how the models work - instead, we use mathematical conversion to approximate their character-level performance based on their predictions.
|
| 111 |
+
|
| 112 |
+
Perplexity fairly evaluates how well each model handles:
|
| 113 |
+
- Spelling accuracy across a diverse vocabulary
|
| 114 |
+
- Grammar rules that span multiple words
|
| 115 |
+
- Sentence structure and flow
|
| 116 |
+
- Language-specific patterns (like how different languages form plurals or compound words)
|
| 117 |
+
**Why does this Matter?** Models with lower perplexity generally perform better on real-world tasks like text generation, translation, and understanding context. It's a reliable indicator of overall language competency across different applications.
|
| 118 |
+
**What data did we use?**
|
| 119 |
+
We use WMT24++ as it is a multilingual, language-parallel evaluation set that none of the models have seen during training. WMT24++ is a composite of texts from news, literature, speech, and social media; thus, it is suitable for foundational model benchmarking.
|
| 120 |
+
| Language | TildeOpen-30B | Gemma-2-27B | EuroLLM-9B | ALIA-40B |
|
| 121 |
+
|----------|---------------|-------------|------------|-----------------|
|
| 122 |
+
| Bulgarian | **2.1716** | 2.3541 | 2.3502 | 2.2411 |
|
| 123 |
+
| Croatian | **2.2259** | 2.6809 | 2.6780 | 2.3456 |
|
| 124 |
+
| Czech | **2.2682** | 2.4873 | 2.4808 | 2.3639 |
|
| 125 |
+
| Danish | **2.0968** | 2.2608 | 2.2586 | 2.1543 |
|
| 126 |
+
| Dutch | **2.0136** | 2.1249 | 2.1185 | 2.0629 |
|
| 127 |
+
| English | 2.1497 | **2.0342** | 2.1897 | 2.1027 |
|
| 128 |
+
| Estonian | **2.2825** | 2.7163 | 2.5652 | 2.4232 |
|
| 129 |
+
| Finnish | **2.1687** | 2.4069 | 2.3844 | 2.2774 |
|
| 130 |
+
| French | 1.9779 | 2.0195 | 2.0479 | **1.9750** |
|
| 131 |
+
| German | **1.9664** | 2.0214 | 2.0499 | 1.9725 |
|
| 132 |
+
| Hungarian | **2.1481** | 2.3308 | 2.3705 | 2.2493 |
|
| 133 |
+
| Icelandic | **2.2011** | 3.1917 | 5.3162 | 4.0978 |
|
| 134 |
+
| Italian | **2.0431** | 2.1065 | 2.1213 | 2.0604 |
|
| 135 |
+
| Latvian | **2.2477** | 2.6701 | 2.4896 | 2.4352 |
|
| 136 |
+
| Lithuanian | **2.2301** | 2.5495 | 2.4754 | 2.4109 |
|
| 137 |
+
| Norwegian | **2.2445** | 2.4173 | 2.5121 | 2.3152 |
|
| 138 |
+
| Polish | **2.1214** | 2.2294 | 2.2264 | 2.1847 |
|
| 139 |
+
| Portuguese | **2.0810** | 2.1554 | 2.1561 | 2.0884 |
|
| 140 |
+
| Romanian | **2.1266** | 2.2724 | 2.2821 | 2.1974 |
|
| 141 |
+
| Russian | **2.1502** | 2.2091 | 2.2813 | 2.1889 |
|
| 142 |
+
| Serbian | **2.3708** | 2.8053 | 4.7160 | 2.5119 |
|
| 143 |
+
| Slovak | **2.2281** | 2.4674 | 2.4588 | 2.3505 |
|
| 144 |
+
| Slovenian | **2.2662** | 2.5798 | 2.5087 | 2.3611 |
|
| 145 |
+
| Spanish | 2.0400 | 2.0665 | 2.1186 | **2.0055** |
|
| 146 |
+
| Swedish | **2.1471** | 2.2971 | 2.2856 | 2.2039 |
|
| 147 |
+
| Turkish | **2.2108** | 2.3665 | 2.3508 | 3.0611 |
|
| 148 |
+
| Ukrainian | **2.2470** | 2.4000 | 2.4251 | 2.3168 |
|
| 149 |
+
|