Commit
·
d397d1a
1
Parent(s):
9b9a502
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,32 @@
|
|
| 1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
tags:
|
| 3 |
+
- pytorch
|
| 4 |
+
- causal-lm
|
| 5 |
+
metrics:
|
| 6 |
+
- accuracy
|
| 7 |
+
language:
|
| 8 |
+
- sl
|
| 9 |
license: apache-2.0
|
| 10 |
---
|
| 11 |
+
|
| 12 |
+
# GPT-sl-base
|
| 13 |
+
|
| 14 |
+
This model is a Slovene GPT model, based on the [bigscience workshop](https://github.com/bigscience-workshop/Megatron-DeepSpeed) fork of the Megatron. GPT-sl-base was trained on large Slovene corpora: Gigafida, KAS, slWaC, and MaCoCu.
|
| 15 |
+
|
| 16 |
+
## Model architecture
|
| 17 |
+
GPT-sl-base has about 110 million parameters. It consists of 12 transformer layers with a dimension of 768. It has 16 attention heads and can process sequences up to 1024 tokens in length.
|
| 18 |
+
The tokenizer was trained on a smaller subset of the corpora, and has the vocabulary of 60k tokens.
|
| 19 |
+
|
| 20 |
+
## Training
|
| 21 |
+
The model was trained for about 20 epochs, a total of 390k steps or 102B tokens seen during training.
|
| 22 |
+
|
| 23 |
+
| Step | Validation Perplexity |
|
| 24 |
+
|:------:|:---------------------:|
|
| 25 |
+
| 50000 | 26.801 |
|
| 26 |
+
| 100000 | 25.574 |
|
| 27 |
+
| 150000 | 24.773 |
|
| 28 |
+
| 200000 | 24.099 |
|
| 29 |
+
| 250000 | 23.336 |
|
| 30 |
+
| 300000 | 22.607 |
|
| 31 |
+
| 350000 | 22.329 |
|
| 32 |
+
| 390000 | 22.293 |
|