Corpus size in Gemma tokens with SP-token note
Browse files
README.md
CHANGED
|
@@ -29,7 +29,8 @@ configs:
|
|
| 29 |
# ArmWeb
|
| 30 |
|
| 31 |
**ArmWeb** is a curated Armenian news corpus for language-model pretraining:
|
| 32 |
-
**4.37M documents /
|
|
|
|
| 33 |
Western Armenian text with document-level provenance (URL, outlet, topic,
|
| 34 |
publication date). To our knowledge it is the first and largest open
|
| 35 |
pretraining corpus *built specifically for* Armenian — existing Armenian text
|
|
|
|
| 29 |
# ArmWeb
|
| 30 |
|
| 31 |
**ArmWeb** is a curated Armenian news corpus for language-model pretraining:
|
| 32 |
+
**4.37M documents / 11 GB / 3.3B Gemma-4 tokens** (1.15B tokens in the 32K
|
| 33 |
+
SentencePiece tokenizer used for the small-scale ablations) of deduplicated, decontaminated Eastern and
|
| 34 |
Western Armenian text with document-level provenance (URL, outlet, topic,
|
| 35 |
publication date). To our knowledge it is the first and largest open
|
| 36 |
pretraining corpus *built specifically for* Armenian — existing Armenian text
|