Update README.md
Browse files
README.md
CHANGED
|
@@ -18,4 +18,47 @@ tags:
|
|
| 18 |
- gpt2
|
| 19 |
- from-scratch
|
| 20 |
- polish-gpt2
|
| 21 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
- gpt2
|
| 19 |
- from-scratch
|
| 20 |
- polish-gpt2
|
| 21 |
+
---
|
| 22 |
+
|
| 23 |
+
## Description
|
| 24 |
+
This is the polish gpt2 model in small architecture.
|
| 25 |
+
This model was released on 11.08.2023
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
## Datasets
|
| 29 |
+
Data which are used to train this model:
|
| 30 |
+
- clarin-knext/msmarco-pl
|
| 31 |
+
- clarin-knext/nq-pl
|
| 32 |
+
- clarin-knext/hotpotqa-pl
|
| 33 |
+
- clarin-knext/scidocs-pl
|
| 34 |
+
- clarin-knext/nfcorpus-pl
|
| 35 |
+
- clarin-knext/dbpedia-pl
|
| 36 |
+
- clarin-knext/trec-covid-pl
|
| 37 |
+
- clarin-knext/quora-pl
|
| 38 |
+
- clarin-knext/arguana-pl
|
| 39 |
+
- clarin-knext/fiqa-pl
|
| 40 |
+
- own corpora not published yet
|
| 41 |
+
|
| 42 |
+
It is about 10,5 GB of data.
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
## Metrics from W&B
|
| 46 |
+
|
| 47 |
+
- train/loss: 2.9569
|
| 48 |
+
- train/train_samples_per_second: 31.797
|
| 49 |
+
- train/epoch: 20
|
| 50 |
+
- train/train_steps_per_second: 3.18
|
| 51 |
+
- train/total_flos: 16645483478384640000
|
| 52 |
+
- train/train_loss: 3.106043342053213
|
| 53 |
+
- train/learning_rate: 2.2070550413783577e-8
|
| 54 |
+
- train/global_step: 3185240
|
| 55 |
+
- train/train_runtime:1001735.8967
|
| 56 |
+
- eval/samples_per_second: 57.896
|
| 57 |
+
- eval/runtime: 1447.4458
|
| 58 |
+
- eval/steps_per_second: 5.79
|
| 59 |
+
- eval/loss: 2.890829086303711
|
| 60 |
+
- eval/accuracy: 0.4637797431547294
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
## Changelog
|
| 64 |
+
- _Nothing_
|