Dimiter Milushev

5/Everything is open source: Dataset: huggingface.co/datasets/dmilu… Model: huggingface.co/dmilush/shield… Code: github.com/dvm81/shieldlm Writeup: @dimiter.milushev/shieldlm-a-unified-dataset-and-classifier-for-prompt-injection-detection-50d5acc513af" target="_blank" rel="nofollow noopener">medium.com/@dimiter.milus…

4/It's not a silver bullet. Jailbreak detection is the weakest at 93.5%. English-dominant. Single-turn only. And at real-world attack rates (<0.1%), even 0.1% FPR means false alarms outnumber real detections. It's Layer 1 in a defense stack — fast pre-filter, not the final word.

1/I built an open-source prompt injection detector. DeBERTa-v3-base trained on a curated dataset of 54K samples. 96.1% detection rate at 0.1% false positive rate — same architecture as ProtectAI v2, but +17pp from better training data.

@capitalcom . Ä

Trade smart with a reliable and stable broker — Capital .com Get everything you need to seize your trading opportunity.

NeRD and Neural-PIL enable decompositions into shape, BRDF, and lighting. However, they require camera poses. Common in-the-wild collections are not only in multiple illuminations but also locations. Here, COLMAP fails. SAMURAI aims to fix this: markboss.me/publication/20… 🧵

🔴 Светът има проблем - руският паранаичният диктатор #Путин започна война срещу #Украйна. Заплашва и всички по #света, които му се противопоставят. Не, не е сценарий на евтин екшън.

@jeremyphoward I am interested in translating the next one i Bulgarian.

Thanks & congrats to Ekaterina Kolodezeva, who just finished translating the final part 1 course.fast.ai video to Russian! She tells me the process really helped her understand the material. Other language contributions most welcome. :) youtube.com/timedtext_vide…

My non-inclusive tweet for the year: I'm looking forward to the time when all of the new year's resolution gym members realize a certain amount of commitment is needed and most stop going. It's far too crowded during the first few weeks of a new year.

No substitute to expert knowledge ... smerity.com/articles/2018/…

Excited to share our #emnlp2018 work: SOTA on many NLP tasks where performances already feel saturated! ideas: models self-train on unlabeled data by refraining access from full representations of text sequences + multitask-traing. by @clark_kev, @chrmanning, @quocleix & I.

@BSidesZurich Many thanks for organizing, I had a great time at my first BSides Zurich. Great job on selecting a really nice mix of interesting speakers who were able to cover a wide spectrum of topics!

Really cool twitter.com/seb_ruder/stat…

NASA just released these 8 images of Jupiter taken by the Juno spacecraft, the highest resolution, most color detailed to date.

"I interviewed at 5 top companies in 5 days, and got 5 job offers" - honest article about interviews & preparation @XiaohanZeng/i-interviewed-at-five-top-companies-in-silicon-valley-in-five-days-and-luckily-got-five-job-offers-25178cf74e0f" target="_blank" rel="nofollow noopener">medium.com/@XiaohanZeng/i…

Really nice talk! twitter.com/sansforensics/…

Merry Christmas everyone!

My plane was delayed, so this is the setup for my #fsharpconf talk in 20 minutes... #hoorayforairportwifi