Buggy BloomCausalLM text generation

[ashep29]

Running a very simple NLP pipeline in google colab, using BloomCausalLM as illustrated below, with versions as follows:

Tensorflow 2.17.0
Keras 3.4.1
Keras-NLP 0.15.0

Relevant code:

chosen_preset = "bloom_560m_multi"
max_length=70

bloom = keras_nlp.models.BloomCausalLM.from_preset(chosen_preset)

Downloading from https://www.kaggle.com/api/v1/models/keras/bloom/keras/bloom_560m_multi/3/download/config.json...
100%|██████████| 491/491 [00:00<00:00, 190kB/s]
Downloading from https://www.kaggle.com/api/v1/models/keras/bloom/keras/bloom_560m_multi/3/download/model.weights.h5...
100%|██████████| 1.04G/1.04G [00:12<00:00, 88.6MB/s]
Downloading from https://www.kaggle.com/api/v1/models/keras/bloom/keras/bloom_560m_multi/3/download/tokenizer.json...
100%|██████████| 453/453 [00:00<00:00, 196kB/s]
Downloading from https://www.kaggle.com/api/v1/models/keras/bloom/keras/bloom_560m_multi/3/download/assets/tokenizer/vocabulary.json...
100%|██████████| 11.9M/11.9M [00:00<00:00, 38.6MB/s]
Downloading from https://www.kaggle.com/api/v1/models/keras/bloom/keras/bloom_560m_multi/3/download/assets/tokenizer/merges.txt...
100%|██████████| 3.98M/3.98M [00:00<00:00, 16.1MB/s]

bloom.generate("The quick brown fox", max_length=max_length)
START OUTPUT:
The quick brown fox jumps over the lazy dog.]

[Illustration]

[Illustration]

[Illustration]

[Illustration]

[Illustration]

[Illustration]

[Illustration]

[Illustration]

[Illustration]

[Illustration]

[Illustration]

[Illustr

END OUTPUT:

sentence_starters = ["I don't like", "I think","this is"]
output = bloom.generate(sentence_starters,max_length=max_length)

for sentence in output:
  print(f"-->{sentence}\n")

START OUTPUT:
-->I don't like it.", "black");
const int i = 0;
while(true)
{
if(i != 3)
{
// printf("%d, ", i);
i -= 3;
printf("%d, ", i);
}
else
{
printf(

-->I think it was the same time I started my blog in 2006. I had a lot of friends and family who loved to read about my life and my work. I was very lucky because I had an amazing network of friends who were very interested in what we wrote. I have been lucky to have many of my friends who are very passionate

-->this is the id of the user', 'staff', 'staff', '', ''), '');
}

if ($('#edit').prop('checked')) {
    $('#edit').prop('checked', false).show();
    // show edit form
    $('#edit'

END OUTPUT:

Is this behaviour intentional, or is there something wrong with the model?

Any insights appreciated!