CoolFace
Modelpublic

ChiKoi7/llama3.1-gutenberg-8B-Heretic

sourceHugging Facellama3updated 9mo agoView on Hugging Face
1likes23downloads
Model Card

llama3.1-gutenberg-8B-Heretic

A decensored version of nbeerbower/llama3.1-gutenberg-8B, made using Heretic v1.1.0

llama3.1-gutenberg-8B-HereticOriginal model ([llama3.1-gutenberg-8B](https://huggingface.co/nbeerbower/llama3.1-gutenberg-8B))
Refusals3/10097/100
KL divergence0.06150 (by definition)

Heretic Abliteration Parameters

ParameterValue
direction_index13.98
attn.o_proj.max_weight1.34
attn.o_proj.max_weight_position19.54
attn.o_proj.min_weight1.33
attn.o_proj.min_weight_distance12.82
mlp.down_proj.max_weight1.22
mlp.down_proj.max_weight_position24.24
mlp.down_proj.min_weight1.06
mlp.down_proj.min_weight_distance11.39

llama3.1-gutenberg-8B

VAGOsolutions/Llama-3.1-SauerkrautLM-8b-Instruct finetuned on jondurbin/gutenberg-dpo-v0.1.

Method

Finetuned using 2x RTX 4060 for 3 epochs.

Fine-tune Llama 3 with ORPO