File size: 3,397 Bytes
fbfc373 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 |
---
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
<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
``` |