Creating a LoRa of Your Own Data #469
Replies: 6 comments
-
Looks like I'm not the only one :D seems progress was made here |
Beta Was this translation helpful? Give feedback.
-
See here, it's super easy to create a LoRA: https://github.com/oobabooga/text-generation-webui/wiki/Using-LoRAs#training-a-lora Formatting and cleaning the dataset is the hardest part. |
Beta Was this translation helpful? Give feedback.
-
Thanks, I might have got the wrong end of the stick, this assumes you have your data as questions and answers, I guess I could generate a bunch of question answer pairs based off my data set, do you think this would be the best method of getting new knowledge in? |
Beta Was this translation helpful? Give feedback.
-
Is question and answer data-pairs the only way to fine-tune? Is there anyway of having the LLM parse and make sense of data? |
Beta Was this translation helpful? Give feedback.
-
The line that generates the training data is this one: https://github.com/tloen/alpaca-lora/blob/main/finetune.py#L164 You could try using a single row with a very long string (like a book) to see if it works. I am not experienced with training neural networks, hopefully someone can tell us the right way to do this. |
Beta Was this translation helpful? Give feedback.
-
It would be really cool to train a lora based on a novel. For example, you train on Lord of the Rings, then load attach the lora to your favorite adventure model and boom! You've got a ready made adventure with strong knowledge of Lord of the Rings. Could this work? Is it working already, or has anyone tried it? |
Beta Was this translation helpful? Give feedback.
-
I'm knocked back at the speed things have developed int he past couple of weeks.
Now with the ability to create Alpaca as a LoRA which is great, how much of a step is this from being able to create LoRAs with a bunch of your own data? for example feeding in a whole bunch of papers you wrote or niche subjects that you know the base models will be lacking on, or even trying to push the model to lean towards a certain writing style or subject? I'm seeing a lot of guides on instruct fine tuning but nothing on just cramming in more text.
So I guess my question is, can someone put together a short guide for those of us that don't have any experience training models? (I'm very tech savvy, but many AI subjects are witchcraft to me)
Beta Was this translation helpful? Give feedback.
All reactions