NemoDori-v0.2-Frankend.2-v1-16.6B

Experimental!
A more upscaled version of NemoDori-v0.2-12B-MN-BT, now at 16.6B.

This is also my first successful(?) fine-tuned model using 500 random rows from dataset Inv/c2-logs-cleaned-deslopped in 70 steps. The reason I used that dataset is... just for testing. What I thought is, if I can replace/fill up those duplicated layers by training it, maybe that makes it better.

NemoDori v0.2 is my best merge model so far, but I'm afraid it's still 12B, not much to improve after merging all kinds of models.
Again, I'm just interested to play with these LLM stuff for awhile. Maybe more version of this will come out later.

As far from my short testing, this model has become a little more strict than the parent model (v0.2).I haven't notice anything major yet.
You can use ST with this preset here. Unfortunately, you can't go wild with this model (from my short tests), sometimes it makes little senses, and sometimes... you will get a reddit link (i'm not kidding).

I didn't have enough time to test it, because it's more pricey without quantization.
I trust @mradermacher to make the quants version of this model. (Thank you so much for making those GGUF on my models ^_^)

And... yeah... Your feedbacks are always welcome. Let me know what's your experience using this model, that would be really appreciated.
Take care everyone.

Merge Method

This model was merged from the following models using the passthrough merge method:

• RozGrov/NemoDori-v0.2-12B-MN-BT

🧩 Configuration

slices:
  - sources:
    - model: RozGrov/NemoDori-v0.2-12B-MN-BT
      layer_range: [0, 8]
  - sources:
    - model: RozGrov/NemoDori-v0.2-12B-MN-BT
      layer_range: [8, 24]
      parameters:
        scale:
          - filter: q_proj
            value: 0.919
          - filter: k_proj
            value: 0.919
          - value: 1.0
  - sources:
    - model: RozGrov/NemoDori-v0.2-12B-MN-BT
      layer_range: [16, 24]
      parameters:
        scale:
          - filter: q_proj
            value: 0.7
          - filter: k_proj
            value: 0.7
          - filter: o_proj
            value: 0.0
          - filter: down_proj
            value: 0.0
          - value: 1.0
  - sources:
    - model: RozGrov/NemoDori-v0.2-12B-MN-BT
      layer_range: [16, 32]
      parameters:
        scale:
          - filter: q_proj
            value: 0.919
          - filter: k_proj
            value: 0.919
          - value: 1.0
  - sources:
    - model: RozGrov/NemoDori-v0.2-12B-MN-BT
      layer_range: [32, 40]
merge_method: passthrough
dtype: bfloat16

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "RozGrov/NemoDori-v0.2-Frankend.2-pre"
messages = [{"role": "user", "content": "What is a large language model?"}]

tokenizer = AutoTokenizer.from_pretrained(model)
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])