Text file for finetuning

Generation Models
#1

I am new in AI modeling. My project is this: create models trained with the complete works of a particular author (let’s say, Aristotle) and then ask questions. I guess that the answers I get inside that model will be like the ones that the author would give.
Now, I already used fine-tuning to create a model, using Plato’s works in txt format in Full text of “Plato Complete Works” (archive.org), but I am not getting satisfactory results. What can I do to achieve my goal? Are my assumptions wrong? Do I have to format the txt in a certain way? Any help or comment will be appreciated.

Moris

#2

Hi Moris!

Do you mind sharing an example of the prompts you’re using? They may need to be restructured to model the formatting of the finetuned data.

If you’re still not seeing the results you’d like, we can help you reformat the file to improve performance.

Best,
Ellie

#3

Hello Ellie!

Thanks for responding. I made two attempts with the same text. In the first one, I only uploaded a txt file, with no blank lines. In the second, I added a ### as a separator between paragraphs. That’s all. The text is Human Action, by L. von Mises (+800 Pages). My aim is to get Mises’ responses to present economic issues.

Best regards,

Moris
Universidad Francisco Marroquin

El El mié, 15 de dic. de 2021 a la(s) 1:53 p. m., ellie via Cohere Community <cohere@discoursemail.com> escribió:

#4

Hi @mpolanco! Thanks for reaching out. Can you share a few screenshots of the inputs you’re using to query the finetuned models? You may have some success rewording your questions or giving the model a few examples of the output you’re looking for- Take a look at our guide on how to engineering your inputs to get the best results from our models. Let us know if this doesn’t solve your problem either.

Cheers!
Elaine

#5

Thank you, Elaine. I made two inputs. The first one was just a plain txt file, Human Action, by L. von Mises; I deleted the empty lines, but that was it. In the second, I put ### between paragraphs of the same book. I guess I am not doing things all right. Can you instruct me?

Best,

Moris

#7

Hello Moris,
We recently discovered a bug with medium model finetuning, which we fixed this morning. If you re-upload your dataset, you should see significant performance improvements.

We apologize for the inconvenience. As a thank you for your patience, we’re sending you one of Cohere’s custom models finetuned on philosophy text. This medium-sized model will be live in your playground shortly.

If you’d like to schedule a 15-minute product demo, don’t hesitate to reach out to support@cohere.ai.

Best,
Ellie

1 Like
#8

Ellie,

Those are good news. Just one question: did you send me the philosophy text, or should I wait for another email?

Take care,

Moris

#9

I would really appreciate if you send me the Cohere custom model finetuned for philosophy :smiley:

#11

Ellie,

I am still waiting for

#12

Hi Moris,
Apologies for the delay on our end. You should see the custom model in your playground now.

Let us know if you have any additional questions.

Best,
Ellie

#13

Thanks a lot. I apologize for my impatience. I see the new model. Can I see the txt you used for finetuning?

Moris

#14

I really need to see the txt .