Nicolas

xpgx1

xpgx1
xpgx1

AI & ML interests

I'm just another LLM enthusiast, in it for the thrill of this new field! Local LLMs and generative AI (Stable Diffusion) are a new era of supercharged creativity for me. I'm merely using the frameworks, enjoying the hard work of all model trainers and project maintainers. Keep on keeping on - we need reliable, unbiased and LOCAL models in all areas!

Recent Activity

liked a model 11 days ago

Virt-io/SillyTavern-Presets

liked a model 11 days ago

DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters

liked a model 18 days ago

zer0int/CLIP-GmP-ViT-L-14

View all activity

Organizations

None yet

xpgx1's activity

liked 2 models 11 days ago

Virt-io/SillyTavern-Presets

Updated Sep 25 • 242

DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters

Updated 5 days ago • 28

liked a model 18 days ago

zer0int/CLIP-GmP-ViT-L-14

Zero-Shot Image Classification • Updated Sep 23 • 181k • 322

liked a model 20 days ago

comfyanonymous/flux_text_encoders

Updated 11 days ago • 613

liked a model 23 days ago

DavidAU/MN-Dark-Horror-The-Cliffhanger-18.5B-GGUF

Text Generation • Updated 18 days ago • 3.55k • 5

liked 3 models 28 days ago

Reacted to m-ric's post with 👀 28 days ago

Post

1541

🧠 𝗖𝗟𝗘𝗔𝗥: 𝗳𝗶𝗿𝘀𝘁 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝘁𝗼 𝗺𝗮𝗸𝗲 𝗺𝗼𝗱𝗲𝗹𝘀 𝗳𝗼𝗿𝗴𝗲𝘁 𝘄𝗵𝗮𝘁 𝘄𝗲 𝘄𝗮𝗻𝘁 𝘁𝗵𝗲𝗺 𝘁𝗼 𝗳𝗼𝗿𝗴𝗲𝘁

With privacy concerns rising, we sometimes need our models to "forget" specific information - like a person's data - while keeping everything else intact. Researchers just released CLEAR, the first benchmark to test how well this works with both text and images.

❌ Bad news: Current methods either fail to truly forget or end up forgetting way too much. It's like trying to remove a single ingredient from a baked cake!

✨ But there's hope: Adding simple mathematical constraints (L1 regularization) during the forgetting process significantly improves results.

🎯 Key insights:

✅ The benchmark tests forgetting on 200 fictional personas
‣ 3,770 visual Q&A pairs
‣ 4,000 textual Q&A pairs
‣ Additional real-world tests

🛑 Most current forgetting methods don't work well with both text and images
‣ They either remember what they should forget
‣ Or they forget too much unrelated information

✨ Simple mathematical constraints work surprisingly well
‣ L1 regularization prevents excessive forgetting
‣ Works especially well with the LLMU method

👉 Read the full paper here: CLEAR: Character Unlearning in Textual and Visual Modalities (2410.18057)