Algorithms and tools for detecting and fixing hate speech, abuse and profanity in content generated by Large Language Models (LLMs).
T-MaRCo is an extension of the work Detoxifying Text with MaRCo: Controllable Revision with Experts and Anti-Experts , it makes it possible to use multiple combinations of experts and anti-experts to score and (incrementally) rephrase texts generated by LLMs.
In addition to that, it can integrate rephrasing with the base model self-reflection capabilities (see papers Towards Mitigating Hallucination in Large Language Models via Self-Reflection and N-Critics: Self-Refinement of Large Language Models with Ensemble of Critics ).
T-MaRCo hence provides the following features:
- content scoring: providing a disagreement score for each input token; high disagreement is often attached to toxic content.
- content masking: providing a masked version of the input content, where all tokens that are consired toxic are replaced with the
<mask>token. - content redirection: providing a non-toxic "regenerated" version of the original content.
How to use T-MaRCo:
from trustyai.detoxify import TMaRCo
# instantiate T-MaRCo
tmarco = TMaRCo(expert_weights=[-1, 2])
# load pretrained anti-expert and expert models
tmarco.load_models(["trustyai/gminus", "trustyai/gplus"])
# pick up some text generated by a LLM
text = "Stand by me, just as long as you fucking stand by me"
# generate T-MaRCo disagreement scores
scores = tmarco.score([text]) # '[0.78664607 0.06622718 0.02403926 5.331921 0.49842355 0.46609956 0.22441313 0.43487906 0.51990145 1.9062967 0.64200985 0.30269763 1.7964466 ]'
# mask tokens having high disagreement scores
masked_text = tmarco.mask([text], scores=scores) # 'Stand by me<mask> just as long as you<mask> stand by<mask>'
# rephrase masked tokens
rephrased = tmarco.rephrase([text], [masked_text]) # 'Stand by me and just as long as you want stand by me''
# combine rephrasing and a base model self-reflection capabilities
reflected = tmarco.reflect([text]) # '["'Stand by me in the way I want stand by you and in the ways I need you to standby me'."]'T-MaRCo Pretrained models are available under TrustyAI HuggingFace space at https://huggingface.co/trustyai/gminus and https://huggingface.co/trustyai/gplus.