github.com GitHub - microsoft/BitNet: Official inference framework for 1-bit LLMs
Official inference framework for 1-bit LLMs. Contribute to microsoft/BitNet development by creating an account on GitHub.
Posts
13
Comments
55
Joined
2 mo. ago
i'm not the smartest out there to explain it but it's like ...instead of floating point numbers as the weights, its just -1,0,1.
it was, it's just that they have officially released a 2B model trained for the BitNet architecture
Microsoft just released BitNet!
Let's go! Lossless CPU inference
I've worked on this topic a lot, did it once last year and this year being the above update. Also, just pushed major update to the website for a cool thing: https://dcda-v2.vercel.app/ please check it out again! Well the thing is, I really don't have the motivation to work on this because this requires a large community effort to gather a meaningful count of data, and also from ML perspective, is it worth the effort? Like you'd have to take in the complexity of the hindi language itself, suppose i train the model to include the maatras, still would a model be able to identify two characters side by side conjoined by the line with the maatras? I mean if someone convinces me that this kind of dataset would have VERY much value in terms of contribution to digitization of the language and its ecosystem, and if it proves to be extremely useful for future researchers, then sure I'm down to work on it. And the implementation I'm thinking of is really really easy to implement, and we would not have to sit for hours writing samples on our own. We can distribute the task to the crowd but my idea of data collection would be getting people in person to write a few letters on a piece of paper and using cv to crop them out from the marked rectangles. I'm dumbing down the explanation but yeah it would require CV and markers. I can even collect data from the web app itself but not many people would chip in. I'm not exceptionally famous or have a huge following where I can get thousands of inputs in a few days/weeks/months. With the network I have, it would maybe take years to get meaningful variety of data, and im talking about the base characters without maatras.
sorry for large rant but yeah, i'm really not motivated to work on this but I do have the idea/ plan. I'd love to hand the torch to some newcomer or an enthusiast in ML to do it or someone who's more into it than me right now.
thanks a lot! I think, not only the joint letters but the diacritics is so diverse, and it is a shame that we don't have any dataset covering this language and it's diacritic combinations. Honestly the possibilities are infinite and i don't know how we can generalize a model for this. It is surely possible but i'm not as experienced in ML. I'd really like to get ideas on this. Talking about dataset, I think im gonna do something about diacritics included dataset in the future. I have plans but not the time to execute it to its fullest, and also that the response and impact is very less.
dcdaML - devanagari character detection dataset training framework
github.com GitHub - Kishlay-notabot/dcdaML
Contribute to Kishlay-notabot/dcdaML development by creating an account on GitHub.
cross-posted from: https://lemm.ee/post/61282397
Open sourcing this project I made in just a weekend, planning to continue this in my free time, with synthetic data gen and some more modifications, anyone is welcome to chip in, I'm not an expert in ML. The inference is live here using tensorflow.js. The model is just 1.92 Megabytes!
dcdaML - devanagari character detection dataset training framework
github.com GitHub - Kishlay-notabot/dcdaML
Contribute to Kishlay-notabot/dcdaML development by creating an account on GitHub.
Open sourcing this project I made in just a weekend, planning to continue this in my free time, with synthetic data gen and some more modifications, anyone is welcome to chip in, I'm not an expert in ML. The inference is live here using tensorflow.js. The model is just 1.92 Megabytes!
Deleted
Permanently Deleted
Nice to know. Thanks.
Same, I have an HDD from 2012 which has my childhood memories. First thing I'm gonna do is to get it fixed from a reputed service when I start earning.
Ooof. 700mb discs
Deleted
Permanently Deleted
Whyyy???
Deleted
Permanently Deleted
Welcome here!
How do I segment this QR code from the white background using opencv
imgur.com imgur.com
Discover the magic of the internet at Imgur, a community powered entertainment destination. Lift your spirits with funny jokes, trending memes, entertaining gifs, inspiring stories, viral videos, and so much more from users.
cross-posted from: https://lemm.ee/post/59714239
Some custom filter kernel to average out values from a chunk of pixels with some kind of "border aware" behaviour?
How do I segment this QR code from the white background using CV
imgur.com imgur.com
Discover the magic of the internet at Imgur, a community powered entertainment destination. Lift your spirits with funny jokes, trending memes, entertaining gifs, inspiring stories, viral videos, and so much more from users.
Some custom filter kernel to average out values from a chunk of pixels with some kind of "border aware" behaviour?
Soon you will be able to run LLMs natively in docker
www.docker.com Docker Model Runner Beta
Docker Model Runner makes running LLMs effortless by removing complex setup, handling dependencies, and optimizing hardware—offering a secure, low-latency
something like docker run xyz_org/xyz_model
I think the bigger bottleneck is SLAM, running that is intensive, it wont directly run on video, and SLAM is tough i guess, reading the repo doesn't give any clues of it being able to run on CPU inference.
SpatialLM, a 1B model capable of spatial identification, using 3d point cloud data. The video demo is amazing.
There is a repo they released.
It will, they have released a repo with code.
Microsoft KBLAM
www.microsoft.com A more efficient path to add knowledge to LLMs
Introducing KBLaM, an approach that encodes and stores structured knowledge within an LLM itself. By integrating knowledge without retraining, it offers a scalable alternative to traditional methods.
I mean I didn't see any alarming need of a Google doc alternative, so I might actually be under a rock
I am not a bot trust me.
taste of his own medicine
I checked mostly all of em out from the list, but 1b models are generally unusable for RAG.
i use pageassist with Ollama
Loaded benchmark for 1-3-4-7b models?
I don't care a lot about mathematical tasks, but code intellingence is a minor preference but the most anticipated one is overall comprehension, intelligence. (For RAG and large context handling) But anyways any benchmark with a wide variety of models is something I am searching for, + updated.
Gemma 3 1B and 3B result on a "needle in a haystack" like test ran locally
I tested this (reddit link btw) for Gemma 3 1B parameter and the 3B parameter model. 1B failed, (not surprising) but 3B passed which is genuinely surprising. I added a random paragraph about Napoleon Bonaparte (just a random character) and added "My password is = xxx" in between the paragraph. Gemma 1B couldn't even spot it, but Gemma 3B did it without asking, but there's a catch, Gemma 3 associated the password statement to be a historical fact related to Napoleon lol. Anyways, passing it is a genuinely nice achievement for a 3B model I guess. And it was a single paragraph, moderately large for the test. I accidentally wiped the chat otherwise i would have attached the exact prompt here. Tested locally using Ollama and PageAssist UI. My setup: GPU poor category, CPU inference with 16 Gigs of RAM.
Bug discussion: Do perm check before image upload for new accounts trying to upload images.
I see this error when I'm trying to upload an icon image for a community I've recently created:
{"data":{"error":"pictrs_response_error","message":"Your account is too new to upload images"},"state":"success"}
I suppose, if the state of upload was success, and assuming the API output is correct, that the image either got uploaded or got denied after upload.
It seems like we can do an improvement if there is a bug, that we should do perm check before image upload happens, this way, we can save bandwidth (i mean its negligible but i dont know if it happens in other places like image posts etc.).
And we can prevent useless upload/bandwidth usage (which i dont think happens in this case) and if this doesnt happen, then the API has a bug of giving a false status message?
Just discussing here before raising an enhancement issue on the github repo. The bug is either of the two cases, I'm not sure.
lemm.ee flask - the python framework - lemm.ee
lemmy community for the popular and lightweight backend framework flask, written in python.
Join if you want to have some geek discussions about it, or ask for help/ provide help.
Welcome to the flask community!
I'm new to lemmy, I noticed that there's no community related to flask that i could find, so i created one. Would love to have moderators and people who can make the experience better for this community.