Hmm. What are the strange characters in your training data, line breaks etc? What format does it take, a specific Json format? Also, is it not possible to have LLama output formatted newsletter? Why the extra step of formatting it with Claude?
So far yes. Can it get better? Yes. This has helped me get drafts to the 70-80% complete mark off the rip though so has helped shorten the time significantly.
Thanks for your reply. Have you found a way to host the fine tuned model somewhere so that you can access it via API? And maybe use it with make/zapier?
@@heymichaeldaigler can you spot some differences between the base model and the one finetuned with alpaca? If so, do you recommend fine-tuning first on alpaca ( even a dataset of another language X like Italian for example ) and then finetuning on our own data ( of the same language X)?
For uploading the dataset and formatting: The file should be a list of JSON objects with the following keys: "Instruction", "Input", "Output". "Instruction"-an example of what a user instruction would look like. "Input"-example of accompanying input with the instruction (if needed) "Output"-example of the format, style, etc of the output we expect --- once you have your dataset file you will: 1. click "new dataset" on your hugging face account 2. Go to files & versions. Select upload file. Upload your dataset. 3. Copy the name of the dataset which will usually be "your_huggingface_username/name_of_dataset" --- lmk if that helps with your issue!
When you say, "make it local" do you mean using newsletters or writing not published online? if so, then you can, as long as it's formatted the right way for the dataset
![[DRIP] ‘CLIK CLAK’ PREVIEW](http://i.ytimg.com/vi/bcxkcOSbpac/mqdefault.jpg)
Hmm. What are the strange characters in your training data, line breaks etc? What format does it take, a specific Json format? Also, is it not possible to have LLama output formatted newsletter? Why the extra step of formatting it with Claude?
are you happy with the results of your fine tuned model? how would you compare it with the article done on claude 3 ( using your ontology method) ?
So far yes. Can it get better? Yes. This has helped me get drafts to the 70-80% complete mark off the rip though so has helped shorten the time significantly.
Thanks for your reply. Have you found a way to host the fine tuned model somewhere so that you can access it via API? And maybe use it with make/zapier?
ok my dataset is private right how will it access
Why did you finetuned the basic model with alpaca first? Couldn’t you just finetune the basic model on your newsletters?
I found that with just fine-tuning it on my newsletters only, gave less optimal results
@@heymichaeldaigler can you spot some differences between the base model and the one finetuned with alpaca?
If so, do you recommend fine-tuning first on alpaca ( even a dataset of another language X like Italian for example ) and then finetuning on our own data ( of the same language X)?
I have a csv of question answer pairs how should I upload it plzzz answer
For uploading the dataset and formatting:
The file should be a list of JSON objects with the following keys: "Instruction", "Input", "Output".
"Instruction"-an example of what a user instruction would look like.
"Input"-example of accompanying input with the instruction (if needed)
"Output"-example of the format, style, etc of the output we expect
---
once you have your dataset file you will:
1. click "new dataset" on your hugging face account
2. Go to files & versions. Select upload file. Upload your dataset.
3. Copy the name of the dataset which will usually be "your_huggingface_username/name_of_dataset"
---
lmk if that helps with your issue!
Can you make it local without using online just dont include the scraping part
When you say, "make it local" do you mean using newsletters or writing not published online? if so, then you can, as long as it's formatted the right way for the dataset
Couldn’t you achieve similar results with an agent?
Yes, you could setup a team of agents with different roles and probably have good results. Just wanted to show/explore other ways to do it.
using python though... kinda counter productive, anyone working with this will have to sit for ages just trying to solve meta questions