Llama3.1-SuperHawk-8B-Heretic

A decensored version of Llama3.1-SuperHawk-8B, made using Heretic v1.1.0

Note: A different version of the model, where all the models in the merge were abliterated using Heretic before merging, can be found here

	Llama3.1-SuperHawk-8B-Heretic	Original model (Llama3.1-SuperHawk-8B)
Refusals	4/100	99/100
KL divergence	0.0949	0 (by definition)

Heretic Abliteration Parameters

Parameter	Value
direction_index	14.37
attn.o_proj.max_weight	1.43
attn.o_proj.max_weight_position	23.02
attn.o_proj.min_weight	1.24
attn.o_proj.min_weight_distance	14.37
mlp.down_proj.max_weight	1.40
mlp.down_proj.max_weight_position	27.22
mlp.down_proj.min_weight	0.75
mlp.down_proj.min_weight_distance	15.05

---

---

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

• Joseph717171/Llama-3.1-SuperNova-8B-Lite_TIES_with_Base
• mukaj/Llama-3.1-Hawkish-8B

🧩 Configuration

slices:
  - sources:
      - model: Joseph717171/Llama-3.1-SuperNova-8B-Lite_TIES_with_Base
        layer_range: [0, 32]
      - model: mukaj/Llama-3.1-Hawkish-8B
        layer_range: [0, 32]
merge_method: slerp
base_model: Joseph717171/Llama-3.1-SuperNova-8B-Lite_TIES_with_Base
parameters:
  t:
    - value: 0.3
dtype: bfloat16

💻 Usage

!pip install -qU transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "Yuma42/Llama3.1-SuperHawk-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.	31.14
IFEval (0-Shot)	79.86
BBH (3-Shot)	31.97
MATH Lvl 5 (4-Shot)	23.49
GPQA (0-shot)	8.39
MuSR (0-shot)	10.38
MMLU-PRO (5-shot)	32.73