Examples of toxic prompts we removed

Announcing RefCOCO-M, a refreshed RefCOCO with pixel-accurate masks and the problematic prompts removed. Better data for better evaluation.

huggingface.co/datasets/moo...

go to moondream.ai/c/playground for description

lone pip user among uv sheeple

matz is nice and so we kick out people starting ridiculous witch hunts

you sound like a markov chain. just parroting random things without any underlying comprehension

wrong. that's exactly how human cognition works

i bet she can't even write down the flash attention 2 backward pass without looking it up

wrong yet highly confident is not a good look

Boars detected!

Are there any boars?

The Moondream VLM (vision language model) is amazing. It's small (2B parameters) and yet accurately detects boars in a dark infrared camera image. Asking for both boar and pig increases accuracy though. #sanglicam

One night the pupil came in tears to Suiwo. "I must return south in shame and embarrassment," he said, "for I cannot solve my problem."

Suiwo said: "Meditate for three days longer, then if you fail to attain enlightenment, you had better kill yourself."

On the second day the pupil was enlightened.

😭

new moondream, new me

huggingface.co/vikhyatk/moo...

Gaze detection will be in the upcoming moondream release!

Live demo: huggingface.co/spaces/moond...
Blog post: moondream.ai/blog/announc...

yeah… in latest torch on h100s it’s basically the same speed

bsky.app/profile/jay....

i heard they have prebuilt wheels but you have to go to the github repo to find them... just stopped using flash attention instead

how much do we actually love the open web?

I noticed that my posts are not in this dataset, may I ask why? Have I offended you somehow?

Moondream playground is back, running the latest upcoming version of the model, and now with prompt suggestions! Look at how blazingly fast it is running on an L40S...

Try it out here: moondream.ai/playground

💀

while we're ruining your weekend here's another one i think you'll like 😂 arxiv.org/pdf/2410.06205

When Precision Meets Position: BFloat16 Breaks Down RoPE in Long-Context Training Extending context window sizes allows large language models (LLMs) to process longer sequences and handle more complex tasks. Rotary Positional Embedding (RoPE) has become the de facto standard due to...

my post was prompted by this paper btw i just realized i didn't share context arxiv.org/abs/2411.13476

is the amazing length extrapolation it enables really worth all of the suffering entailed in debugging when it breaks? 🥲

when are we going to admit RoPE was a mistake?

working on screen understanding

talking to customers is overrated