base_model, tags, license, language, pipeline_tag
| base_model | tags | license | language | pipeline_tag | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
|
other |
|
text-generation |
NeutralWeirdness v1 24B Heretic
A merge of WeirdCompound v1.7 and NeutralGear V.2, then abliterated using Heretic for uncensored output.
Description of This Model
This is my first published model. I merged the WeirdCompound v1.7 and NeutralGear V.2 via a 50/50 SLERP merge, then the refusal direction was removed using Heretic v1.2.0.
I havent tested the model too much. but it has a good stable and creative role-playing capability.
Merge Recipe
- Method: SLERP 50/50
- Tool: mergekit
- Model A: FlareRebellion/WeirdCompound-v1.7-24b
- Model B: OddTheGreat/NeutralGear_24B_V.2
Abliteration
- Tool: Heretic v1.2.0
- Trial used: Trial 161 (100/100 Refusals to 8/100 refusals, KL divergence 0.0471)
- Quantization during process: bnb_4bit
- Device: NVIDIA RTX 3090 24GB
Recommended Settings
- Template: Mistral V7 Tekken
- Sampler: top-p 0.8, min-p: 0.1 temp 0.75
- Context: up to 32k
Hardware Used
- Merge: Ryzen 5 3600, 32GB RAM (CPU only)
- Abliteration: RTX 3090 24GB
Credits
- mradermacher for Quantization
- FlareRebellion for WeirdCompound
- OddTheGreat for NeutralGear
- FailSpy for abliterator
- p-e-w for Heretic
- arcee-ai for mergekit