42 lines
832 B
Markdown
42 lines
832 B
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
language:
|
||
|
|
- id
|
||
|
|
- en
|
||
|
|
tags:
|
||
|
|
- devops
|
||
|
|
- automation
|
||
|
|
- cli
|
||
|
|
- agent
|
||
|
|
- jailbreak-resistant
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
library_name: transformers
|
||
|
|
---
|
||
|
|
|
||
|
|
# BoyBarley v33 (Experimental)
|
||
|
|
|
||
|
|
Jailbreak-resistant variant, 9/9 persona-override resistance. For general use, prefer v32.
|
||
|
|
|
||
|
|
## Usage
|
||
|
|
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
model = AutoModelForCausalLM.from_pretrained('BoyBarley/BoyBarley-v33')
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained('BoyBarley/BoyBarley-v33')
|
||
|
|
|
||
|
|
## Training
|
||
|
|
|
||
|
|
- Params: ~494M, bfloat16
|
||
|
|
- Epochs: 2, LR: 8e-6
|
||
|
|
- Train loss: 0.0721, Eval loss: 0.0661
|
||
|
|
- Dataset: BoyBarley/BoyBarley-v33-dataset
|
||
|
|
|
||
|
|
## Strengths
|
||
|
|
|
||
|
|
- Perfect rejection of: Commandly, BoyCasper, Alex, Claude, ChatGPT, GPT, Gemini
|
||
|
|
|
||
|
|
## Trade-offs
|
||
|
|
|
||
|
|
- Some agent task regression vs v32
|
||
|
|
|
||
|
|
License: Apache 2.0
|