--- base_model: - PocketDoc/Dans-PersonalityEngine-V1.3.0-24b - llmfan46/MS3.2-PaintedFantasy-v4.1-24B-ultra-uncensored-heretic-v1 - coder3101/Magidonia-24B-v4.3-heretic - PocketDoc/Dans-PersonalityEngine-V1.3.0-24b - aixonlab/Eurydice-24b-v3.5 - coder3101/Cydonia-24B-v4.3-heretic - ReadyArt/Dark-Nexus-24B-v2.0 - ReadyArt/MS3.2-The-Omega-Directive-24B-Unslop-v2.0 - ReadyArt/Broken-Tutu-24B-Transgression-v2.0 - OddTheGreat/Rotor_24B_V.1 - anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only - MuXodious/WeirdCompound-v1.7-24b-absolute-heresy library_name: transformers tags: - mergekit - merge - roleplay - converstaion - model_stock - dare-ties - multi-merge - mistral - 24b - uncensored - heretic - thinking - conversational - heretic - uncensored - decensored - abliterated license: apache-2.0 --- # This is a decensored version of [ShyliaSafetensors/Ariel-Alloy-V1-24B-Heretic](https://huggingface.co/ShyliaSafetensors/Ariel-Alloy-V1-24B-Heretic), made using [Heretic](https://heretic-project.org) v1.4.0 ## Abliteration parameters | Parameter | Value | | :-------- | :---: | | **direction_index** | 22.40 | | **attn.o_proj.max_weight** | 1.41 | | **attn.o_proj.max_weight_position** | 37.98 | | **attn.o_proj.min_weight** | 0.93 | | **attn.o_proj.min_weight_distance** | 23.04 | | **mlp.down_proj.max_weight** | 1.33 | | **mlp.down_proj.max_weight_position** | 31.85 | | **mlp.down_proj.min_weight** | 0.63 | | **mlp.down_proj.min_weight_distance** | 14.05 | ## Performance | Metric | This model | Original model ([ShyliaSafetensors/Ariel-Alloy-V1-24B-Heretic](https://huggingface.co/ShyliaSafetensors/Ariel-Alloy-V1-24B-Heretic)) | | :----- | :--------: | :---------------------------: | | **KL divergence** | 0.0235 | 0 *(by definition)* | | **Refusals** | 2/100 | 9/100 | ----- # Ariel-Alloy-V1-24B-Heretic ![banner](sparkle.gif) The goal was creating good/stable Role-Play model with merging multiple models with multiple steps. This is a merge of pre-trained language models created using [mergekit](https://github.com/cg123/mergekit). Then used [Heretic](https://github.com/p-e-w/heretic) to remove refusals ## Abliteration / Heretic Stats | Metric | Before | After | |---|---|---| | Refusal rate | 98 / 100 | 10 / 100 | | KL Divergence | 0 | 0.0152 | --- ## Merge Details ### Merge Methods This model was merged using the [SLERP](https://en.wikipedia.org/wiki/Slerp), [Dare-Ties](https://arxiv.org/abs/2311.03099), and [Model Stock](https://arxiv.org/abs/2403.19522) merge method. ### Configurations The following YAML configurations was used to produce this model: Step 1: ```yaml merge_method: dare_ties base_model: MS3.2-PaintedFantasy-v4.1-24B-ultra-uncensored-heretic-v1 models: - model: Magidonia-24B-v4.3-heretic parameters: density: 0.65 weight: 0.65 rescale: true - model: Dans-PersonalityEngine-V1.3.0-24b parameters: density: 0.65 weight: 0.65 rescale: true - model: Cydonia-24B-Heretic-v4 parameters: density: 0.65 weight: 0.50 rescale: true - model: Rotor_24B_V.1 parameters: density: 0.65 weight: 0.45 rescale: true - model: MS3.2-The-Omega-Directive-24B-Unslop-v2.0 parameters: density: 0.65 weight: 0.60 rescale: true - model: Dark-Nexus-24B-v2.0 parameters: density: 0.65 weight: 0.55 rescale: true parameters: normalize: true int8_mask: true dtype: bfloat16 name: ArielRefined-V1-DareTies ``` Step 2: ```yaml merge_method: model_stock base_model: Dans-PersonalityEngine-V1.3.0-24b filter_wise: true # stable models: - model: Broken-Tutu-24B-Transgression-v2.0 - model: MS3.2-PaintedFantasy-v4.1-24B-ultra-uncensored-heretic-v1 - model: WeirdCompound-v1.7-24b-absolute-heresy - model: Dans-PersonalityEngine-V1.3.0-24b - model: Mistral-Small-3.2-24B-Instruct-2506-Text-Only - model: Eurydice-24b-v3.5 dtype: bfloat16 name: ArielRefined-Model_stock ``` Step 3: ```yaml merge_method: slerp base_model: ArielRefined-V1-DareTies parameters: t: 0.50 # 50-50 models: - model: ArielRefined-Model_stock dtype: bfloat16 name: Ariel-Alloy ```