Add preset for finetuning GPT2 on CNN news

[chenmoneygithub]

With the finetuning, the generated text stops properly at end token, and the writing is better.

To play with the preset, install this branch:

!pip install -q -U git+https://github.com/chenmoneygithub/keras-nlp.git@add-gpt2-news

then use the code below:

import keras_nlp

gpt2_lm = keras_nlp.models.GPT2CausalLM.from_preset("gpt2_base_en_cnn_dailymail")
gpt2_lm.generate(
    ["that's weird", "that's even weirder"],
    sampler="top_k",
    max_length=100,
)

[mattdangerw]

Nice this seems useful to have. Just a few comments.

[mattdangerw]

I think the most similar checkpoint we have here is the sst2 dataset for bert. So maybe we should stick to the more precise naming and call this, gpt2_base_en_cnn_dailymail?

[chenmoneygithub]

sure!

[mattdangerw]

Approving! Let's update our file paths before we merge.

[mattdangerw]

Update URLs to the new path before we merge this. Also remember to add a /v1/ to the path as above, so we can update our checkpoints in the future.

[chenmoneygithub]

done!