|
--- |
|
language: |
|
- en |
|
metrics: |
|
- accuracy |
|
co2_eq_emissions: |
|
emissions: "10" |
|
source: "mlco2.github.io" |
|
training_type: "fine-tuning" |
|
geographical_location: "West Java, Indonesia" |
|
hardware_used: "1 T4" |
|
license: cc-by-nc-sa-4.0 |
|
widget: |
|
- text: 'You: "Hey kekbot! Whats up?"\nKekbot: "' |
|
example_title: "Asking what's up" |
|
- text: 'You: "Hey kekbot! How r u?"\nKekbot: "' |
|
example_title: "Asking how he is" |
|
--- |
|
> THIS MODEL IS INTENDED FOR RESEARCH PURPOSES ONLY |
|
# Kekbot Mini |
|
|
|
Based on a `distilgpt2` model, fine-tuned to a select subset (65k<= messages) of Art Union's general-chat channel chat history. |
|
|
|
### Limits and biases |
|
As this is trained on chat history, it is possible that discriminatory or even offensive materials to be outputted. |
|
Author holds his ground on the fact that ML models are mere statistical representation of the dataset used to train it, |
|
and that due to the nature of the dataset it is practically impossible to be certain of |
|
the degree of "cleanliness" that the data contained within holds. |
|
|
|
Author can confirm, however, that from heuristical testing that the model was not found to be offensive |
|
to the author himself, hopefully this opinion stays true for everyone in the audience. |
|
|
