This is an uncensored version of
Llama 3.1 70B Instruct created with abliteration (see
this article to know more about it) using
@grimjim's recipe.
I adapted this recipe to Llama 3.1 70B using
failspy/Meta-Llama-3-70B-Instruct-abliterated-v3.5 and optimized the LoRA rank.
The model is fully uncensored in my tests and maintains a high level of quality. A more rigorous evaluation is still needed to measure the impact of this process on benchmarks.
Special thanks to
@grimjim for this technique (see his
8B model) and
@FailSpy for his
70B abliterated model. Please follow them if you're interested in abliterated models.
In addition, thanks to
brev.dev for providing me with compute!
General-purpose, role-play (see feedback from
McUH). Use the Llama 3 chat template.
This model was merged using the
task arithmetic merge method using ./meta-llama/Meta-Llama-3.1-70B-Instruct + Llama-3-70B-Instruct-abliterated-LORA as a base.
1base_model: meta-llama/Meta-Llama-3.1-70B-Instruct+Llama-3-70B-Instruct-abliterated-LORA
2dtype: bfloat16
3merge_method: task_arithmetic
4parameters:
5 normalize: false
6slices:
7- sources:
8 - layer_range: [0, 80]
9 model: meta-llama/Meta-Llama-3.1-70B-Instruct+Llama-3-70B-Instruct-abliterated-LORA
10 parameters:
11 weight: 1.0
1# Setup
2git clone https://github.com/arcee-ai/mergekit.git
3cd mergekit && pip install -e .
4pip install bitsandbytes
5
6# Extraction
7mergekit-extract-lora failspy/Meta-Llama-3-70B-Instruct-abliterated-v3.5 meta-llama/Meta-Llama-3-70B-Instruct Llama-3-70B-Instruct-abliterated-LORA --rank=64
8
9# Merge using previous config
10mergekit-yaml config.yaml Llama-3.1-70B-Instruct-lorablated --allow-crimes --lora-merge-cache=./cache
