L3.2-3B-Herthea

A DARE merge combining abliterated instruction-tuning with RP capabilities for Llama 3.2 3B.

Merge Details

This model merges:

• DavidAU/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored (base)
• lunahr/thea-rp-3b-25r (weighted at 0.9)

Created using Steelskull/mergekit-gui.

Merge Method

DARE (Drop And REscale) merge with TIES resolution, (hopefully) combining an abliterated base with strong RP tuning.

Configuration

models:
  - model: DavidAU/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored
  - model: lunahr/thea-rp-3b-25r
    parameters:
      weight: 0.9
      density: 0.5
merge_method: dare_ties
base_model: DavidAU/Llama-3.2-3B-Instruct-heretic-ablitered-uncensored
dtype: bfloat16

Usage

Important: When using GGUF quantizations in inference engines like KoboldCpp, manually set the chat template to Llama 3 format. Auto-detection may fail and cause degraded output quality or formatting issues.

Llama 3.2 uses the standard Llama 3 chat format:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>
{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>
{user_message}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Characteristics

• Reduced refusals from abliterated base.
• Good roleplay and creative writing capabilities.
• Character consistency and narrative coherence.
• Instruction following from base model (should be) preserved.

Quantizations

You can make your own GGUF quants via GGUF-my-repo.

Notes

• Designed for creative writing and roleplay applications.
• Should have reduced safety guardrails due to abliteration.
• Best results with proper system prompts and chat formatting.
• Not intended for factual or safety-critical use cases, obviously.

Acknowledgments

Merge created using the mergekit-gui space by Steelskull, which made this merge possible without local hardware requirements. (I'm limited to a phone)