Ivan Mikhnenkov PRO

ivanmikhnenkov

1 1 7

https://ivanmikhnenkov.com/

AI & ML interests

Computer Vision, Gen AI, LLMs, Reinforcement Learning

Recent Activity

liked a model about 12 hours ago

SupraLabs/Supra2-IMG

reacted to anakin87's post with ❤️ 15 days ago

I made a 1.1M ModernBERT encoder play Doom in real time on a CPU Some time ago, VAGO Solutions released SauerkrautLM-Doom-MultiVec-1.3M, a tiny model trained to play Doom Defend the Center scenario from 31k human gameplay examples. My first thought: cool! I love both Doom and Small Language Models. Then another idea: I bet I can do better :-) What I did? - evaluated the original model and found it's better than reported - changed a bit the architecture - generated SFT data with a scripted oracle - SFT + PPO refinement on consumer hardware Got a smaller, faster and killer model Can even fit a floppy with int8 quantization 💾 Watch it play/read the article: https://huggingface.co/spaces/anakin87/tiny-doom-defender

new activity 15 days ago

ivanmikhnenkov/tinydit-256:SupraLabs join request

View all activity

Organizations

liked a model about 12 hours ago

SupraLabs/Supra2-IMG

Text-to-Image • Updated 4 days ago • 378 • 117

reacted to anakin87's post with ❤️ 15 days ago

Post

3453

I made a 1.1M ModernBERT encoder play Doom in real time on a CPU

Some time ago, VAGO Solutions released SauerkrautLM-Doom-MultiVec-1.3M, a tiny model trained to play Doom Defend the Center scenario from 31k human gameplay examples.

My first thought: cool! I love both Doom and Small Language Models.

Then another idea: I bet I can do better :-)

What I did?
- evaluated the original model and found it's better than reported
- changed a bit the architecture
- generated SFT data with a scripted oracle
- SFT + PPO refinement on consumer hardware

Got a smaller, faster and killer model
Can even fit a floppy with int8 quantization 💾

Watch it play/read the article: anakin87/tiny-doom-defender

New activity in ivanmikhnenkov/tinydit-256 15 days ago

SupraLabs join request

🔥 2

#1 opened 16 days ago by

LH-Tech-AI

published an article 15 days ago

Article

Trained 210M text-to-image model from scratch on one GPU: what actually mattered

ivanmikhnenkov

•

15 days ago

• 10

posted an update 16 days ago

Post

126

I trained a 210M text-to-image diffusion transformer from scratch on one GPU in 3.5 days, on 4.2M curated images at 256². Same six prompts and seeds at every checkpoint of the run 👇

🎨 Try it: ivanmikhnenkov/tinydit

What mattered most: captions that fit the images, a timestep shift for the 32-channel FLUX.2 latent, aspect-ratio buckets from step one, register tokens with learned null attention slots, and torch.compile for training.

🤗 Weights: ivanmikhnenkov/tinydit-256
💻 Code, every decision with its sources: https://github.com/ivanmikhnenkov/tinydit