File size: 2,396 Bytes
647ca25 fdb74d2 647ca25 fdb74d2 dab1798 21a75b2 647ca25 02a4bad b28f7ec 02a4bad b28f7ec 02a4bad b28f7ec 02a4bad b28f7ec 02a4bad b28f7ec f6fc929 b28f7ec 02a4bad b28f7ec 02a4bad b28f7ec |
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 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 |
---
datasets:
- wiki-corpus
language:
- lt
license: apache-2.0
tags:
- "text-generation"
widget:
- text: "Lietuvos yra viena "
---
## Model description
GPT-2 model from Lithuania using Wikipedia corpus dataset based on GPT-2 small model.
This is only the first version of the model, over time model will be improved using a bigger dataset and better data preparation.
## Training data
This model was pre-trained with 180MB of Lithuanian Wikipedia. The texts are tokenized using a byte-level version of Byte Pair Encoding (BPE).
## Training
The model was trained on wiki-corpus for 40 hours using NVIDIA Tesla P100 GPU.
## How to use
### Load model
``` from transformers import AutoTokenizer, TFAutoModelWithLMHead
import tensorflow as tf
tokenizer = AutoTokenizer.from_pretrained("DeividasM/gpt2_lithuanian_small")
model = TFAutoModelWithLMHead.from_pretrained("DeividasM/gpt2_lithuanian_small")
# Get sequence length max of 1024
tokenizer.model_max_length=1024
model.eval()
```
## Generate text
``` text = "tekstas"
inputs = tokenizer.encode(text, return_tensors="tf")
outputs = model.generate(inputs, eos_token_id=50256, pad_token_id=50256,
do_sample=True,
max_length=40,
top_k=40)
print(tokenizer.decode(outputs[0]))
```
## Limitations and bias
The training data used for this model come from Lithuanian Wikipedia. We know it contains a lot of unfiltered content from the internet, which is far from neutral. As the openAI team themselves point out in their model card:
"Because large-scale language models like GPT-2 do not distinguish fact from fiction, we don’t support use-cases that require the generated text to be true. Additionally, language models like GPT-2 reflect the biases inherent to the systems they were trained on, so we do not recommend that they be deployed into systems that interact with humans > unless the deployers first carry out a study of biases relevant to the intended use-case. We found no statistically significant difference in gender, race, and religious bias probes between 774M and 1.5B, implying all versions of GPT-2 should be approached with similar levels of caution around use cases that are sensitive to biases around human attributes."
## Author
Lithuanian GPT-2 small was trained and evaluated by Deividas Mataciunas
|