A merge/clone of Yuma42/Llama3.1-IgneousIguana-8B but all the models were abliterated before merging using Heretic v1.1.0
To be clear, this repo is NOT the original IgneousIguana-8B model passed through Heretic.
Since the output of this merge had suitably low refusals and a very low KL divergence when evaluated against the original IgneousIguana model, I currently have no version uploaded that IS simply the original model passed through Heretic.
The same merge config as the original model was used, with the only difference being the models were abliterated first.

Note: The output of the merge was evaluated as having 3/100 refusals. There was no need for another pass through Heretic.

	Llama3.1-IgneousIguana-8B-Heretic	Original model (Llama3.1-IgneousIguana-8B)
Refusals	3/100	99/100
KL divergence	0.0822	0 (by definition)

Models Used

ChiKoi7/Llama3.1-SuperHawk-8B-Heretic
Initial Refusals: 99/100
Heretic: 4/100 @ 0.0949 KL Div.

ChiKoi7/llama3.1-gutenberg-8B-Heretic
Initial Refusals: 97/100
Heretic: 4/100 @ 0.0615 KL Div.

Merge Config

slices:
  - sources:
      - model: ChiKoi7/Llama3.1-SuperHawk-8B-Heretic
        layer_range: [0, 32]
      - model: ChiKoi7/llama3.1-gutenberg-8B-Heretic
        layer_range: [0, 32]
merge_method: slerp
base_model: ChiKoi7/Llama3.1-SuperHawk-8B-Heretic
parameters:
  t:
    - value: 0.3
dtype: bfloat16

Llama3.1-IgneousIguana-8B

Llama3.1-IgneousIguana-8B is a merge of the following models using LazyMergekit:

🧩 Configuration

yaml

slices:
  - sources:
      - model: Yuma42/Llama3.1-SuperHawk-8B
        layer_range: [0, 32]
      - model: nbeerbower/llama3.1-gutenberg-8B
        layer_range: [0, 32]
merge_method: slerp
base_model: Yuma42/Llama3.1-SuperHawk-8B
parameters:
  t:
    - value: 0.2
dtype: bfloat16

💻 Usage

python

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "Yuma42/Llama3.1-IgneousIguana-8B"
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"])

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	31.48
IFEval (0-Shot)	81.33
BBH (3-Shot)	31.99
MATH Lvl 5 (4-Shot)	21.98
GPQA (0-shot)	8.05
MuSR (0-shot)	12.47
MMLU-PRO (5-shot)	33.04

ChiKoi7/Llama3.1-IgneousIguana-8B-Heretic

Llama3.1-IgneousIguana-8B-Heretic