🐶 @SentientAGI Dobby's Exploration Journey, Part 5 🐶 Mianjiang and Duanniao teacher @wanghebbf took Dobby to the rooftop and saw glowing envelopes flying in the sky. They used golden string to catch a lost letter. In the old clock tower, Mianjiang used her camera bag to repair the letter, and then a sky full of shooting stars fell... 🌠 This time with Duanniao teacher, I was really romantic, hahahahaha 〰️〰️〰️〰️ Last night, @SentientAGI released OML 1.0 at @NeurIPSConf. After studying their "Extensible LLM Fingerprinting Technology" paper, I found that this solution truly solves a core pain point of open source models: how to maintain openness while protecting model ownership. ▪️Traditional model fingerprinting can only embed up to 100 fingerprints; any more will conflict with each other, significantly impacting model performance. ▪️However, OML 1.0 squeezes 24,576 fingerprints into Llama-3.1-8B with virtually no performance loss. The key lies in their Perinucleus sampling technique—generating fingerprints from the "edge" of the model's probability distribution, which is both natural and conflict-free. Such fingerprints are resistant to: supervised fine-tuning and style training, model distillation and fusion, joint attacks, and key search. The verification method is challenge-response: given a secret key, the model returns a bounded response; without the key, it behaves normally. Best of all, the more fingerprints embedded, the higher the probability of detection, meaning large-scale audits can be performed in real-world applications. 〰️〰️〰️〰️ Okay, all of the above was copied using AI. It's so esoteric and technical that I don't understand it. Anyway, it's awesome! 🤔