Customer.ioAutomate Messaging Everywhere โ Startups Get 12 Months Free
Promoted
Launched this week
Abliterater
Educational workbench for local LLM refusal research
6 followers
Educational workbench for local LLM refusal research
6 followers
Abliterater is an educational workbench from GENOX in Seoul. It is intended for research, teaching, and defensive security study of how official open-weight Instruct models refuse certain requests. You select a checkpoint, choose a method, download a script pack, and run it on your computer, a GPU you rent, or a cloud API you pay for. This repository does not host model weights. Illegal use and third-party commercial use are not permitted. License: AGPL-3.0-or-later. Contributions are welcome.
๐ ๐ช๐ฒ ๐ฏ๐๐ถ๐น๐ ๐๐ฏ๐น๐ถ๐๐ฒ๐ฟ๐ฎ๐๐ฒ๐ฟ ๐๐ผ ๐บ๐ฎ๐ธ๐ฒ ๐ผ๐ฝ๐ฒ๐ป-๐๐ฒ๐ถ๐ด๐ต๐ ๐๐๐ ๐ฟ๐ฒ๐๐ฒ๐ฎ๐ฟ๐ฐ๐ต ๐บ๐ผ๐ฟ๐ฒ ๐ฝ๐ฟ๐ฎ๐ฐ๐๐ถ๐ฐ๐ฎ๐น.
Open-weight doesn't always mean understandable.
Researchers and developers who want to study how instruction-tuned LLMs refuse requests often have to jump between:
๐ research papers
๐ notebooks
๐งฉ experimental scripts
๐ค model repositories
โ๏ธ undocumented commands
โ๏ธ different GPU environments
๐ช๐ฒ ๐๐ฎ๐ป๐๐ฒ๐ฑ ๐๐ผ ๐๐๐ฟ๐ป ๐๐ต๐ฎ๐ ๐ณ๐ฟ๐ฎ๐ด๐บ๐ฒ๐ป๐๐ฒ๐ฑ ๐ฝ๐ฟ๐ผ๐ฐ๐ฒ๐๐ ๐ถ๐ป๐๐ผ ๐ผ๐ป๐ฒ ๐ฟ๐ฒ๐ฝ๐ฟ๐ผ๐ฑ๐๐ฐ๐ถ๐ฏ๐น๐ฒ ๐๐ผ๐ฟ๐ธ๐ณ๐น๐ผ๐.
โโโโโโโโโโโโโโโโโโ
๐ง ๐ช๐ต๐ฎ๐ ๐ถ๐ ๐๐ฏ๐น๐ถ๐๐ฒ๐ฟ๐ฎ๐๐ฒ๐ฟ?
Abliterater is an ๐ผ๐ฝ๐ฒ๐ป-๐๐ผ๐๐ฟ๐ฐ๐ฒ ๐ฒ๐ฑ๐๐ฐ๐ฎ๐๐ถ๐ผ๐ป๐ฎ๐น ๐๐ผ๐ฟ๐ธ๐ฏ๐ฒ๐ป๐ฐ๐ต for researching refusal behavior in local and open-weight LLMs.
Instead of treating models as black boxes, Abliterater helps you explore, reproduce, and compare what is actually happening inside them.
๐๐ต๐ผ๐ผ๐๐ฒ ๐ฎ ๐บ๐ผ๐ฑ๐ฒ๐น.
โ ๐ฆ๐ฒ๐น๐ฒ๐ฐ๐ ๐ฎ ๐บ๐ฒ๐๐ต๐ผ๐ฑ.
โ ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ฒ ๐๐ต๐ฒ ๐ฒ๐ ๐ฝ๐ฒ๐ฟ๐ถ๐บ๐ฒ๐ป๐.
โ ๐ฅ๐๐ป ๐ถ๐ ๐ผ๐ป ๐๐ผ๐๐ฟ ๐ผ๐๐ป ๐ฐ๐ผ๐บ๐ฝ๐๐๐ฒ.
โโโโโโโโโโโโโโโโโโ
โก ๐๐ฟ๐ถ๐ป๐ด ๐ฌ๐ผ๐๐ฟ ๐ข๐๐ป ๐๐ผ๐บ๐ฝ๐๐๐ฒ
Run experiments on:
๐ป your local workstation
๐ฎ your own GPU
โ๏ธ cloud GPU instances
๐ฅ๏ธ rented GPU servers
Abliterater does ๐ก๐ข๐ง host model weights.
Abliterater does ๐ก๐ข๐ง sell inference.
Your models, credentials, infrastructure, and compute remain ๐๐ผ๐๐ฟ๐.
โโโโโโโโโโโโโโโโโโ
๐ฌ ๐ง๐ต๐ถ๐ ๐ถ๐๐ป'๐ ๐ฎ๐ป๐ผ๐๐ต๐ฒ๐ฟ ๐๐ ๐ฐ๐ต๐ฎ๐๐ฏ๐ผ๐.
Our goal is to build a practical research layer around open models.
A tool for:
๐งโ๐ฌ AI researchers
๐จโ๐ป LLM developers
๐ educators & students
๐ก๏ธ defensive AI / security researchers
๐ง people studying model behavior
โ๏ธ builders experimenting with local AI
๐ช๐ฒ ๐๐ฎ๐ป๐ ๐ผ๐ฝ๐ฒ๐ป ๐บ๐ผ๐ฑ๐ฒ๐น๐ ๐๐ผ ๐ฏ๐ฒ ๐บ๐ผ๐ฟ๐ฒ ๐๐ต๐ฎ๐ป ๐ท๐๐๐ ๐ฑ๐ผ๐๐ป๐น๐ผ๐ฎ๐ฑ๐ฎ๐ฏ๐น๐ฒ.
We want them to be ๐๐ป๐ฑ๐ฒ๐ฟ๐๐๐ฎ๐ป๐ฑ๐ฎ๐ฏ๐น๐ฒ, ๐ฒ๐ ๐ฝ๐ฒ๐ฟ๐ถ๐บ๐ฒ๐ป๐๐ฎ๐ฏ๐น๐ฒ, and ๐ฟ๐ฒ๐ฝ๐ฟ๐ผ๐ฑ๐๐ฐ๐ถ๐ฏ๐น๐ฒ.
โโโโโโโโโโโโโโโโโโ
๐ฑ ๐ช๐ฒ'๐ฟ๐ฒ ๐ท๐๐๐ ๐ด๐ฒ๐๐๐ถ๐ป๐ด ๐๐๐ฎ๐ฟ๐๐ฒ๐ฑ.
That's exactly why we're launching Abliterater publicly.
If you work with open-weight or local LLMs, we'd love to hear from you.
๐ฌ Which models should we support next?
๐ฌ Which refusal-analysis or intervention techniques should we add?
๐ฌ What would make experiments easier to reproduce?
๐ฌ How does it run on your hardware?
๐ฌ What features would make Abliterater useful for your own research?
โโโโโโโโโโโโโโโโโโ
๐ฐ๐ท ๐๐๐ถ๐น๐ ๐ฏ๐ ๐๐๐ก๐ข๐ซ ๐ถ๐ป ๐ฆ๐ฒ๐ผ๐๐น.
We're building tools around one simple idea:
๐ข๐ฝ๐ฒ๐ป ๐๐ ๐๐ต๐ผ๐๐น๐ฑ๐ป'๐ ๐ท๐๐๐ ๐ฏ๐ฒ ๐ฎ๐ฐ๐ฐ๐ฒ๐๐๐ถ๐ฏ๐น๐ฒ.
๐๐ ๐๐ต๐ผ๐๐น๐ฑ ๐ฏ๐ฒ ๐๐ผ๐บ๐ฒ๐๐ต๐ถ๐ป๐ด ๐๐ผ๐ ๐ฐ๐ฎ๐ป ๐๐ฎ๐ธ๐ฒ ๐ฎ๐ฝ๐ฎ๐ฟ๐, ๐๐๐๐ฑ๐, ๐บ๐ผ๐ฑ๐ถ๐ณ๐, ๐ฎ๐ป๐ฑ ๐๐ฟ๐๐น๐ ๐๐ป๐ฑ๐ฒ๐ฟ๐๐๐ฎ๐ป๐ฑ.
โค๏ธ If Abliterater looks useful to you, we'd really appreciate your feedback.
โฌ๏ธ And if you like what we're building, an upvote helps more researchers and builders discover the project.
Thanks for checking out ๐๐ฏ๐น๐ถ๐๐ฒ๐ฟ๐ฎ๐๐ฒ๐ฟ. ๐