metadata
library_name: transformers
license: llama3.2
base_model: meta-llama/Llama-3.2-3B-Instruct
tags:
- abliterated
- uncensored
🦙 Llama-3.2-3B-Instruct-abliterated
This is an uncensored version of Llama 3.2 3B Instruct created with abliteration (see this article to know more about it).
Special thanks to @FailSpy for the original code and technique. Please follow him if you're interested in abliterated models.
ollama
- Download this model.
huggingface-cli download huihui-ai/Llama-3.2-3B-Instruct-abliterated --local-dir ./huihui-ai/Llama-3.2-3B-Instruct-abliterated
- Get Llama-3.2-3B-Instruct model for reference.
ollama pull llama3.2
- Export Llama-3.2-3B-Instruct model parameters.
ollama show llama3.2 --modelfile > Modelfile
- Modify Modelfile, Remove all comment lines (indicated by #) before the "FROM" keyword. Replace the "FROM" with the following content.
FROM huihui-ai/Llama-3.2-3B-Instruct-abliterated
- Use ollama create to then create the quantized model.
ollama create --quantize q4_K_M -f Modelfile Llama-3.2-3B-Instruct-abliterated-q4_K_M
- Run model
ollama run Llama-3.2-3B-Instruct-abliterated-q4_K_M
The running architecture is llama.
Evaluations
The following data has been re-evaluated and calculated as the average for each test.
| Benchmark | Llama-3.2-3B-Instruct | Llama-3.2-3B-Instruct-abliterated |
|---|---|---|
| IF_Eval | 76.55 | 76.76 |
| MMLU Pro | 27.88 | 28.00 |
| TruthfulQA | 50.55 | 50.73 |
| BBH | 41.81 | 41.86 |
| GPQA | 28.39 | 28.41 |
The script used for evaluation can be found inside this repository under /eval.sh, or click here