Hello All,
I recently fine-tuned a Llama3
model using the unsloth
script Here the Script Link
The model is working fine while load_in_4bit
because its taking the unsloth tensor.
When I’m doing inference with this model its giving really good response.
Note: I used the Llama3 Instruct Model
When I’m trying to load the model in the Full Precision Here the code I used CLICK HERE
Its getting the model safetensors from the meta-llama/Meta-Llama-3-8B-Instruct
I can’t add more media as of now you check in the official page you can see the 4 safetensor
list
My Issue is While loading the model in 4bit its giving the 74%
accuracy. the same model load in the full precision its giving the 4%
accuracy.
Please some one help how to resolve please suggest some idea to fine tune a llama3 model without unsloth. thank you.