File size: 1,886 Bytes
5cd5870
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ba6be74
5cd5870
ba6be74
 
 
 
5cd5870
ba6be74
 
 
 
 
 
 
 
5cd5870
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
---
license: mit
language:
- en
---

# hmByT5 - Preliminary Language Models

Preliminary Historic Multilingual and Monolingual ByT5 Models. Following languages are currently covered:

* English (British Library Corpus - Books)

More details can be found in [our GitHub repository](https://github.com/stefan-it/hmByT5).

# Pretraining

We use the official JAX/FLAX example in Hugging Face Transformers to pretrain a ByT5 model on a single v3-8 TPU.
Details about the training can be found [here](https://github.com/stefan-it/hmByT5/tree/main/hmbyt5-flax).

This model was trained with `mean_noise_span_length=20` for one epoch.

# Mean Noise Span Length

The previously pretrained hmByT5 models "accidentally" use a mean noise span length of 3, because this value is the
default one for T5. But the ByT5 paper mentions, that using a length of 3 would make pretraining tasks too easy, and
recommend a value of 20. Thus, we pretrained this model with `mean_noise_span_length=20` and fine-tuned it on English
AjMC dataset:

| Configuration                            | Run 1 | Run 2 | Run 3 | Run 4 | Run 5 | Avg.         |
|------------------------------------------|-------|-------|-------|-------|-------|--------------|
| `wsFalse-bs4-e10-lr0.00015-poolingfirst` | 85.48 | 84.6  | 85.65 | 86.83 | 86.53 | 85.82 ± 0.79 |
| `wsFalse-bs4-e10-lr0.00016-poolingfirst` | 85.35 | 84.5  | 86.05 | 85.1  | 85.18 | 85.24 ± 0.5  |
| `wsFalse-bs8-e10-lr0.00016-poolingfirst` | 84.14 | 83.45 | 84.4  | 84.9  | 85.82 | 84.54 ± 0.79 |
| `wsFalse-bs8-e10-lr0.00015-poolingfirst` | 85.27 | 85.3  | 83.33 | 85.25 | 81.7  | 84.17 ± 1.45 |

For comparison the model using a length of 3 achieved 85.65 ± 1.21.

# Acknowledgements

Research supported with Cloud TPUs from Google's [TPU Research Cloud](https://sites.research.google/trc/about/) (TRC).
Many Thanks for providing access to the TPUs ❤️