Fine-tuned model
Thank you for this fantastic model.
I used this model to do fine-tuning with orca mini datasets. Checkout https://huggingface.co/asyafiqe/Merak-7B-v3-Mini-Orca-Indo!
It would be great if you can give some feedback.
Thanks!
Thank you, I have tested it and it works great. :) Now, I also know that Orca is a good dataset for my model.
BTW, since Merak-7B is licensed by CC-BY-SA-NC, that means all derivative works of Merak must be same as Merak which is non commercial & share-alike.
I will try fine tune Merak-v3 with other dataset to make it better.
Thanks again for contributing in LLM world :)
Great!
I see. Thank you for the PR. I've already merged it.
Can't wait to see your next fine-tuning!
Great!
I see. Thank you for the PR. I've already merged it.
Can't wait to see your next fine-tuning!
Thank you for your understanding.. :)
BTW, I have a problem that I still find the solution. Both Merak and Merak-Orca occur hallucination with history question.
For example : "Siapa komposer dari lagu kebangsaan Indonesia Raya?"
My original Merak has been fine tuned by Wikipedia. I sure I have trained that. But, The hallucination still occurs.
My next fine tune, maybe I train again Merak with Wikipedia dataset. It will use another hyperparameter or maybe another prompt. So it hopefully reduce the hallucination.
I think that is the problem of fine-tuning. Fine tuning cannot directly add knowledge, it's more of adjusting response style. The knowledge depends more on the pretraining. It is often said that for accurate answer, retrieval augmented generation (RAG) is more preferable.
Anyway, Merak v3 already consistently answers in Bahasa Indonesia. Response is also much longer than v2.
I think that is the problem of fine-tuning. Fine tuning cannot directly add knowledge, it's more of adjusting response style. The knowledge depends more on the pretraining. It is often said that for accurate answer, retrieval augmented generation (RAG) is more preferable.
I see... I still find any solution since Indonesian language is only 0,03% in the pretraining dataset (according to the Llama2 paper).
RAG and RHLF is next possible solution. I will learn it too.
Anyway, Merak v3 already consistently answers in Bahasa Indonesia. Response is also much longer than v2.
Praising to the God, I'm happy because v3 has response which is much longer than v2.
I think it was because I have trained it with Ichsan2895/OASST_Top1_Indonesian & Ichsan2895/alpaca-gpt4-indonesian