metadata

base_model:
  - NousResearch/Hermes-3-Llama-3.1-70B
  - Fizzarolli/L3.1-70b-glitz-v0.2
  - cyberagent/Llama-3.1-70B-Japanese-Instruct-2407
  - Sao10K/L3-70B-Euryale-v2.1
tags:
  - merge
  - mergekit
  - lazymergekit
  - NousResearch/Hermes-3-Llama-3.1-70B
  - Fizzarolli/L3.1-70b-glitz-v0.2
  - cyberagent/Llama-3.1-70B-Japanese-Instruct-2407
  - Sao10K/L3-70B-Euryale-v2.1

L3.1-70b-Ginny

L3.1-70b-Ginny is a merge of the following models using LazyMergekit:

I really liked Glitz and Euryale. Though they can get kinda wish-washy and don't follow structure well enough. I used Hermes as a base as it has rather good instruct following but it's way too instruct focused. I find myself running into Japanese text too. Which neither 3 models are like superb at, so I used cyberagent's Japanese Instruct to give it a boost.

🧩 Configuration


models:
  - model: NousResearch/Hermes-3-Llama-3.1-70B
    parameters:
      density: 0.33
      weight: 0.25
  - model: Fizzarolli/L3.1-70b-glitz-v0.2
    parameters:
      density: 0.7
      weight: 0.5
  - model: cyberagent/Llama-3.1-70B-Japanese-Instruct-2407
    parameters:
      density: 0.5
      weight: 0.25
  - model: Sao10K/L3-70B-Euryale-v2.1
    parameters:
      density: 0.7
      weight: 0.5

merge_method: ties
base_model: NousResearch/Hermes-3-Llama-3.1-70B
parameters:
  normalize: true
dtype: bfloat16

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "KaraKaraWitch/L3.1-70b-Ginny"
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"])