|
## BabyBERTA |
|
|
|
### Overview |
|
|
|
BabyBERTa is a light-weight version of RoBERTa trained on 5M words of American-English child-directed input. |
|
It is intended for language acquisition research, on a single desktop with a single GPU - no high-performance computing infrastructure needed. |
|
|
|
## Loading the tokenizer |
|
|
|
BabyBERTa was trained with `add_prefix_space=True`, so it will not work properly with the tokenizer defaults. |
|
Make sure to load the tokenizer as follows: |
|
|
|
```python |
|
tokenizer = RobertaTokenizerFast.from_pretrained("phueb/BabyBERTa", |
|
add_prefix_space=True) |
|
``` |
|
|
|
### Performance |
|
|
|
The provided model is the best-performing out of 10 that were evaluated on the [Zorro](https://github.com/phueb/Zorro) test suite. |
|
This model was trained for 400K steps, and achieves an overall accuracy of 80.3, |
|
comparable to RoBERTa-base, which achieves an overall accuracy of 82.6 on the latest version of Zorro (as of October, 2021). |
|
|
|
Both values differ slightly from those reported in the paper (Huebner et al., 2020). |
|
There are two reasons for this: |
|
1. Performance of RoBERTa-base is slightly larger because the authors previously lower-cased all words in Zorro before evaluation. |
|
Lower-casing of proper nouns is detrimental to RoBERTa-base because RoBERTa-base has likely been trained on proper nouns that are primarily title-cased. |
|
In contrast, because BabyBERTa is not case-sensitive, its performance is not influenced by this change. |
|
2. The latest version of Zorro no longer contains ambiguous content words such as "Spanish" which can be both a noun and an adjective. |
|
this resulted in a small reduction in the performance of BabyBERTa. |
|
|
|
|
|
### Additional Information |
|
|
|
This model was trained by [Philip Huebner](https://philhuebner.com), currently at the [UIUC Language and Learning Lab](http://www.learninglanguagelab.org). |
|
|
|
More info can be found [here](https://github.com/phueb/BabyBERTa). |