Merge of MiniMerlin via Task arithmetic using mergekit. There was no goal except merging. Interest in the outcome tho. I might need to fine-tune it more.

FT on more french data (Merlin).

Je pense qu'il s'agit du meilleur model français en 3B. Essayez le.

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

model = AutoModelForCausalLM.from_pretrained(
    "teilomillet/Potato-3B",
    revision="0.1",
    return_dict=True,
    torch_dtype=torch.bfloat16,
    device_map='auto'
)

tokenizer = AutoTokenizer.from_pretrained("teilomillet/Potato-3B")
tokenizer.pad_token = tokenizer.eos_token

text = "[|User|] Comment faire un bon plat ? </s>[|Assistant|]"
inputs = tokenizer(text, return_tensors="pt").to(0)

outputs = model.generate(**inputs, max_new_tokens=800)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))

#merge

RichardErkhov/teilomillet_-_Potato-3B-4bits

Original model description:

RichardErkhov/
teilomillet_-_Potato-3B-4bits