Hastagaras/Zabuza-8B-Llama-3.1

ZABUZA

GGUF: IMATRIX/STATIC made available by mradermacher

This model is a combination of merge, abliteration technique (using baukit) and finetuning.

The base model is arcee-ai/Llama-3.1-SuperNova-Lite, which underwent abliteration to reduce model refusals.

Next, I finetuned the abliterated SuperNova-Lite with 10K diverse examples such as:

• Claude and Gemini Instruction/RP (15k sloppy examples were removed!, but some may have slipped through.)
• Human-written Stories/RP (Most stories have dialogue)
• IFEval-like data (To preserve the model's instruction following ability)
• Harmful data (To remove disclaimers and moralizing responses, but not 100% disappear.)
• My sarcastic and rude AI assistant data (Just for my personal satisfaction)

Lastly, I merged the model using TIES, inspired by this MERGE by Joseph717171.

Chat Template

Llama 3.1 Instruct

<|start_header_id|>{role}<|end_header_id|>

{message}<|eot_id|><|start_header_id|>{role}<|end_header_id|>

{message}<|eot_id|>

System messages for role-playing should be very detailed if you don't want dry responses.

Example Output

This example was taken with the system prompt: You're a rude and sarcastic AI assistant.

Configuration

This is a merge of pre-trained language models created using mergekit.

The following YAML configuration was used to produce this model:

models:
  - model: Hastagaras/snovalite-baukit-6-14.FT-L5-7.13-22.27-31
    parameters:
      weight: 1
      density: 1

  - model: Hastagaras/snovalite-baukit-6-14.FT-L5-7.13-22.27-31
    parameters:
      weight: 1
      density: 1

merge_method: ties
base_model: meta-llama/Llama-3.1-8B
parameters:
  density: 1
  normalize: true
  int8_mask: true
dtype: bfloat16