15 82 357

alkinun

AtAndDev

AI & ML interests

LLMs, Alignment, Merging, Unsloth, DPO, SFT, ORPO, SPIN..

Recent Activity

reacted to merve's post with 🤗 about 11 hours ago

So many open releases at Hugging Face past week 🤯 recapping all here ⤵️ https://huggingface.co/collections/merve/march-21-releases-67dbe10e185f199e656140ae 👀 Multimodal > Mistral AI released a 24B vision LM, both base and instruction FT versions, sota 🔥 (OS) > with IBM we released SmolDocling, a sota 256M document parser with Apache 2.0 license (OS) > SpatialLM is a new vision LM that outputs 3D bounding boxes, comes with 0.5B (QwenVL based) and 1B (Llama based) variants > SkyWork released SkyWork-R1V-38B, new vision reasoning model (OS) 💬 LLMs > NVIDIA released new Nemotron models in 49B and 8B with their post-training dataset > LG released EXAONE, new reasoning models in 2.4B, 7.8B and 32B > Dataset: Glaive AI released a new reasoning dataset of 22M+ examples > Dataset: NVIDIA released new helpfulness dataset HelpSteer3 > Dataset: OpenManusRL is a new agent dataset based on ReAct framework (OS) > Open-R1 team released OlympicCoder, new competitive coder model in 7B and 32B > Dataset: GeneralThought-430K is a new reasoning dataset (OS) 🖼️ Image Generation/Computer Vision > Roboflow released RF-DETR, new real-time sota object detector (OS) 🔥 > YOLOE is a new real-time zero-shot object detector with text and visual prompts 🥹 > Stability AI released Stable Virtual Camera, a new novel view synthesis model > Tencent released Hunyuan3D-2mini, new small and fast 3D asset generation model > ByteDance released InfiniteYou, new realistic photo generation model > StarVector is a new 8B model that generates svg from images > FlexWorld is a new model that expands 3D views (OS) 🎤 Audio > Sesame released CSM-1B new speech generation model (OS) 🤖 Robotics > NVIDIA released GR00T, new robotics model for generalized reasoning and skills, along with the dataset *OS ones have Apache 2.0 or MIT license

reacted to Jaward's post with 🔥 about 12 hours ago

Implemented a custom multimodal GRPO trainer that scales for Small VLMs, supports cpu and gpu with vllm + flash attention. Using SmolVLM-256M-Instruct reference & reward model, wasn’t trained for long btw, still got some sparks of “thinking”:) Code: https://github.com/Jaykef/ai-algorithms/blob/main/grpo_multimodal_reasoner.ipynb

replied to MonsterMMORPG's post 3 days ago

Extending Wan 2.1 generated video - First 14b 720p text to video, then using last frame automatically to to generate a video with 14b 720p image to video - with RIFE 32 FPS 10 second 1280x720p video Our app has this fully automated : https://www.patreon.com/posts/123105403 Here how it works image : https://ibb.co/b582z3R6 Workflow is easy Use your favorite app to generate initial video. Get last frame Give last frame to image to video model - with matching model and resolution Generate And merge Then use MMAudio to add sound I made it automated in my Wan 2.1 app but can be made with ComfyUI easily as well . I can extend as many as times i want :) Here initial video Prompt: Close-up shot of a Roman gladiator, wearing a leather loincloth and armored gloves, standing confidently with a determined expression, holding a sword and shield. The lighting highlights his muscular build and the textures of his worn armor. Negative Prompt: Overexposure, static, blurred details, subtitles, paintings, pictures, still, overall gray, worst quality, low quality, JPEG compression residue, ugly, mutilated, redundant fingers, poorly painted hands, poorly painted faces, deformed, disfigured, deformed limbs, fused fingers, cluttered background, three legs, a lot of people in the background, upside down Used Model: WAN 2.1 14B Text-to-Video Number of Inference Steps: 20 CFG Scale: 6 Sigma Shift: 10 Seed: 224866642 Number of Frames: 81 Denoising Strength: N/A LoRA Model: None TeaCache Enabled: True TeaCache L1 Threshold: 0.15 TeaCache Model ID: Wan2.1-T2V-14B Precision: BF16 Auto Crop: Enabled Final Resolution: 1280x720 Generation Duration: 770.66 seconds

View all activity

Organizations

AtAndDev's activity

reacted to merve's post with 🤗 about 11 hours ago

Post

1753

So many open releases at Hugging Face past week 🤯 recapping all here ⤵️ merve/march-21-releases-67dbe10e185f199e656140ae

👀 Multimodal
> Mistral AI released a 24B vision LM, both base and instruction FT versions, sota 🔥 (OS)
> with IBM we released SmolDocling, a sota 256M document parser with Apache 2.0 license (OS)
> SpatialLM is a new vision LM that outputs 3D bounding boxes, comes with 0.5B (QwenVL based) and 1B (Llama based) variants
> SkyWork released SkyWork-R1V-38B, new vision reasoning model (OS)

💬 LLMs
> NVIDIA released new Nemotron models in 49B and 8B with their post-training dataset
> LG released EXAONE, new reasoning models in 2.4B, 7.8B and 32B
> Dataset: Glaive AI released a new reasoning dataset of 22M+ examples
> Dataset: NVIDIA released new helpfulness dataset HelpSteer3
> Dataset: OpenManusRL is a new agent dataset based on ReAct framework (OS)
> Open-R1 team released OlympicCoder, new competitive coder model in 7B and 32B
> Dataset: GeneralThought-430K is a new reasoning dataset (OS)

🖼️ Image Generation/Computer Vision
> Roboflow released RF-DETR, new real-time sota object detector (OS) 🔥
> YOLOE is a new real-time zero-shot object detector with text and visual prompts 🥹
> Stability AI released Stable Virtual Camera, a new novel view synthesis model
> Tencent released Hunyuan3D-2mini, new small and fast 3D asset generation model
> ByteDance released InfiniteYou, new realistic photo generation model
> StarVector is a new 8B model that generates svg from images
> FlexWorld is a new model that expands 3D views (OS)

🎤 Audio
> Sesame released CSM-1B new speech generation model (OS)

🤖 Robotics
> NVIDIA released GR00T, new robotics model for generalized reasoning and skills, along with the dataset

*OS ones have Apache 2.0 or MIT license

reacted to Jaward's post with 🔥 about 12 hours ago

Post

1245

Implemented a custom multimodal GRPO trainer that scales for Small VLMs, supports cpu and gpu with vllm + flash attention. Using SmolVLM-256M-Instruct reference & reward model, wasn’t trained for long btw, still got some sparks of “thinking”:)
Code: https://github.com/Jaykef/ai-algorithms/blob/main/grpo_multimodal_reasoner.ipynb

1 reply

replied to MonsterMMORPG's post 3 days ago

I didnt mean you lul
I meant him @MonsterMMORPG ...

replied to MonsterMMORPG's post 4 days ago

brother, dunking on some great models to defend your "product" is not a great (hate to say it but) human value...

replied to nroggendorff's post 4 days ago

ma guys suffered ik :)

reacted to onekq's post with 😔 4 days ago

Post

1512

I like to benchmark 💵o1-pro💵 but it is way too expensive for me 🤦‍♂️

4 replies

replied to onekq's post 4 days ago

Its expensive for everyone, just go with o3-mini, they just figured out that they are not the single llm provider and just doubled the cost of r1 for o3-mini.

reacted to etemiz's post with 🚀😎👀 4 days ago

Post

1644

Started fine tuning Gemma 3 using evolutionary approach. It is not the worst model according to AHA leaderboard and it is one of the smart according to lmarena.ai. My objective is to make it based, anti woke, wise, beneficial and then some.

Several GPUs are fine tuning it at the same time, each using a different dataset and using QLoRA and the successful ones are merged later. Compared to LoRa this allows faster training and also reduced overfitting because the merge operation heals overfitting. The problem with this could be the 4 bit quantization may make models dumber. But I am not looking for sheer IQ. Too much mind is a problem anyway :)

Has anyone tried parallel QLoRa and merge before?

I also automated the dataset selection and benchmarking and converging to objectives (the fit function, the reward). It is basically trying to get higher score in AHA Leaderboard as fast as possible with a diverse set of organisms that "evolve by training".

I want to release some cool stuff when I have the time:
- how an answer to a single question changes over time, with each training round or day
- a chart to show AHA alignment over training rounds

3 replies

liked a model 4 days ago

manycore-research/SpatialLM-Llama-1B

Text Generation • Updated 3 days ago • 1.81k • 559

reacted to clem's post with 👀 4 days ago

Post

3383

Should we assemble affordable open-source robots at Hugging Face for the community. Would you buy them? At what price?

8 replies

upvoted a paper 4 days ago

DeepPerception: Advancing R1-like Cognitive Visual Perception in MLLMs for Knowledge-Intensive Visual Grounding

Paper • 2503.12797 • Published 7 days ago • 27

liked a model 4 days ago

ritvik77/Medical_Doctor_AI_LoRA-Mistral-7B-Instruct_FullModel

Text Generation • Updated 14 days ago • 669 • 6

reacted to clem's post with 🔥 4 days ago

Post

2327

Nice new space to see how fast your personal or organization followers are growing on HF:
julien-c/follow-history

As you can see, I still have more followers than @julien-c even if he's trying to change this by building such cool spaces 😝😝😝

upvoted a collection 4 days ago

Gemma 2 Release

Collection

15 items • Updated 12 days ago • 217

updated a dataset 5 days ago

AtAndDev/ranky-dataset

Viewer • Updated 5 days ago • 2.86k • 70

upvoted a paper 5 days ago

Rewards Are Enough for Fast Photo-Realistic Text-to-image Generation

Paper • 2503.13070 • Published 7 days ago • 9

replied to their post 5 days ago

Also, the links are just wrong as of my knowledge, open source just means its accessible to everyone to download... But the license differs like said, but the worst it can be is not to be used to make money, thats just it.
Please correct me if im wrong.

replied to their post 5 days ago

Well, the models are research and there is some real work going into them but I checked some of those products that are promoted here and they are either clones of spaces you can find here and some name added...
Plus, all models here are oss but licensed different like (cc-by-nc or custom licenses) but either way they provide competition, contribution and ideas here which is always plus to everyone.