README.md

---
library_name: transformers
license: llama3.1
base_model: meta-llama/Meta-Llama-3.1-70B-Instruct
tags:
- abliterated
- uncensored
- mergekit
---

# 🦙 Llama-3.1-70B-Instruct-lorablated

![](https://i.imgur.com/5Y0Riis.png)

<center>🦙 <a href="https://huggingface.co/mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated"><i>Llama 3.1 8B Instruct abliterated</i></a></center>

This is an uncensored version of [Llama 3.1 70B Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-70B-Instruct) created with abliteration (see [this article](https://huggingface.co/blog/mlabonne/abliteration) to know more about it) using [@grimjim](https://huggingface.co/grimjim)'s recipe.

More precisely, this is a **LoRA-abliterated** (lorablated) model:

1. **Extraction**: We extract a LoRA adapter by comparing two models: a censored Llama 3 and an abliterated Llama 3
2. **Merge**: We merge this new LoRA adapter using [task arithmetic](https://arxiv.org/abs/2212.04089) to a censored Llama 3.1 to abliterate it.

I adapted this recipe to Llama 3.1 70B using [failspy/Meta-Llama-3-70B-Instruct-abliterated-v3.5](https://huggingface.co/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](https://huggingface.co/grimjim) for this technique (see his [8B model](https://huggingface.co/grimjim/Llama-3.1-8B-Instruct-abliterated_via_adapter)) and [@FailSpy](https://huggingface.co/failspy) for his [70B abliterated model](https://huggingface.co/failspy/Meta-Llama-3-70B-Instruct-abliterated-v3.5). Please follow them if you're interested in abliterated models.

In addition, thanks to [brev.dev](https://brev.dev/) for providing me with compute!

## 🔍 Applications

General-purpose, role-play (see feedback from [McUH](https://huggingface.co/mlabonne/Llama-3.1-70B-Instruct-lorablated/discussions/7)). Use the Llama 3 chat template.

## ⚡️ Quantization

* **GGUF**: https://huggingface.co/mlabonne/Llama-3.1-70B-Instruct-lorablated-GGUF
* **Bartowski**: https://huggingface.co/bartowski/Llama-3.1-70B-Instruct-lorablated-GGUF (with IQ quants)

## 🧩 Configuration

This model was merged using the [task arithmetic](https://arxiv.org/abs/2212.04089) merge method using ./meta-llama/Meta-Llama-3.1-70B-Instruct + Llama-3-70B-Instruct-abliterated-LORA as a base.

The following YAML configuration was used to produce this model:

```yaml
base_model: meta-llama/Meta-Llama-3.1-70B-Instruct+Llama-3-70B-Instruct-abliterated-LORA
dtype: bfloat16
merge_method: task_arithmetic
parameters:
  normalize: false
slices:
- sources:
  - layer_range: [0, 80]
    model: meta-llama/Meta-Llama-3.1-70B-Instruct+Llama-3-70B-Instruct-abliterated-LORA
    parameters:
      weight: 1.0
```

You can reproduce this model using the following commands:

```bash
# Setup
git clone https://github.com/arcee-ai/mergekit.git
cd mergekit && pip install -e .
pip install bitsandbytes

# Extraction
mergekit-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

# Merge using previous config
mergekit-yaml config.yaml Llama-3.1-70B-Instruct-lorablated --allow-crimes --lora-merge-cache=./cache
```