This is a decensored version of SteelStorage/L3.1-MS-Astoria-70b-v2, made using Heretic v1.1.0

Abliteration parameters

Parameter	Value
direction_index	67.05
attn.o_proj.max_weight	1.40
attn.o_proj.max_weight_position	47.63
attn.o_proj.min_weight	1.04
attn.o_proj.min_weight_distance	13.15
mlp.down_proj.max_weight	1.40
mlp.down_proj.max_weight_position	66.24
mlp.down_proj.min_weight	1.24
mlp.down_proj.min_weight_distance	43.14

Performance

Metric	This model	Original model (SteelStorage/L3.1-MS-Astoria-70b-v2)
KL divergence	0.0268	0 (by definition)
Refusals	11/100	93/100

Now the cute anime girl has your attention

Creator: SteelSkull

About Astoria-70b-v2:

Name Legend:
L3.1 = Llama 3.1
MS = Model Stock
70B = its 70B

This model is a remake of the original astoria with modern models and context sizes its goal is to merge the robust storytelling of mutiple models while attempting to maintain intelligence.

Use Llama 3 Format or meth format (llama 3 refuses to work with stepped thinking but meth works)

Quants: (List of badasses)

GGUF Quant:

- bartowski: Combined-GGUF

- mradermacher: GGUF // Imat-GGUF

Config:

MODEL_NAME = "L3.1-MS-Astoria-70b-v2"
base_model: mlabonne/Llama-3.1-70B-Instruct-lorablated
merge_method: model_stock
dtype: bfloat16
models:
  - model: migtissera/Tess-3-Llama-3.1-70B
  - model: NeverSleep/Lumimaid-v0.2-70B
  - model: Sao10K/L3.1-70B-Euryale-v2.2
  - model: ArliAI/Llama-3.1-70B-ArliAI-RPMax-v1.2
  - model: nbeerbower/Llama3.1-Gutenberg-Doppel-70B

If you wish to support:

sh0ck0r/L3.1-MS-Astoria-70b-v2-heretic