Electric Sheep 7B - α - "The Author"
- Developed by: maldv
- License: cc-by-nc-4.0
- Finetuned from model: maldv/winter-garden-7b-alpha
- Methodology: Simple newline delimited, rolling window book and conversation data.
Will It Write
I spent the first evening after baking this model watching as it wrote page after page, story after story. The answer is, yes. It will write, and never stop. 100% story, 0% plot.
It is one of the most beautiful things I've ever seen. laughs
Data
70% book data, then 10% from each of the other datasets, lora r 64, lr .00007, 2 epochs. Trained for around 2 days on an a6000; loss fell to .4, but that grad leveled out at around 3 and ground in really nicely.
Chat Template
It was trained to follow no prompt at all, just to start going. It can be encouraged by [WP] Topic\n\n, and once it gets going an author note doesn't seem to get regurgitated.
If you have to follow a chat template, use the one it shipped with, as that is what the conversation turns were conditioned on. Format is not super important, but if you use one, then provide a few turns of dialogue to get it rolling.
{{bos_token}}
{% for message in messages %}
{% if 'name' in message %}{{message['name'] + ('' if 'to' not in message else ' (to ' + message['to'] + ')') + ': ' + message['content'] + '\n\n'}}{% else %}{{message['content'] + '\n\n '}}
{% endif %}
{% endfor %}
Which produces results like:
Paul (to Jane): The sky is blue.
Jane (to Paul): Yes, it sure is!