Fine-tuning Babbage with Divine Comedy

I made an experiment in which I fine-tuned a babbage model to reproduce a Divine Comedy’s tercet. The input format is: “[line 1]”
The generate output is: “[line 2]”, “[line 3]”
The fine-tuned works qyute well, the only issue is that the model doesn’t put ‘\n’ at the end of each line, so I have to manually separate them. This is due to an error I did during the JSONL file preparation. The model seems to understand Italian quite well. It didn’t learn to reproduce the rhyme scheme ABA, but it usually generate lines that end with the same letter.

If you want to try it, this is the link of the preset: OpenAI API

2 Likes

Oh, I didn’t know. Then, nobody can’t try this fine-tuned model, even if I provided the link of the preset in the post?

I’ll provide some example here:

nel mezzo del cammin di nostra vita [line 1]
tal cosa non si vede, che non mi si mostri [line 2]
se non a ’l patto che ’l ciel ne fa tra ’l monte. [line 3]

e quando vidi quella fiera feroce [line 1]
che cane e cane affanno, e che li annera [line 2]
le fïate, e li artigli le natiche. [line 3]

e così chiesi alla mia solenne guida [line 1]
‘Come s’a l’etterno consiglio riede’, [line 2]
ch’elli a me rispuose in questo modo. [line 3]

e dal luogo dove sorge la sfera lucente [line 1]
che mira il cielo e sopra le stelle [line 2]
lume a l’infinito vedere avete. [line 3]