Radostin Cholakov (@radi

Radostin Cholakov

594 posts

Radostin Cholakov

@radi_cho

CS @Stanford

Stanford, California

Joined March 2017

Pinned
Radostin Cholakov
@radi_cho
Jun 19
Excited to release cho-embedding-0.8b - a Pareto frontier 0.8B vision-language embedder competitive with some of the most widely-used models up to 4x larger. Built on Qwen3.5-0.8B and post-trained with multi-stage contrastive + distillation objectives, the model achieves 60.7 on
00:00
4K
Radostin Cholakov
@radi_cho
Feb 18, 2023
My latest blog post showcases a minimalistic approach for training text generation architectures from @huggingface with @TensorFlow and Keras as the backend. I am utilizing @Google's mT5 model in a use case from a @kaggle competition. Read it here: medium.com/@radicho/fine-…
39K
Radostin Cholakov
@radi_cho
Mar 16, 2023
Building an open-source command line tool for generating datasets with LLMs. It could be used either to conduct research about the LLM itself (analyze texts, train detectors) or to leverage powerful models (e.g., GPT-3.5) for tuning smaller ones. github.com/radi-cho/datas…
3.8K
Radostin Cholakov
@radi_cho
Nov 11, 2022
Starting today I am officially part of the @GoogleDevExpert program as a Machine Learning GDE! This motivated me to begin tweeting daily about AI/ML and tech in general. Hope more people will be impacted with time.
GIF
Radostin Cholakov
@radi_cho
Sep 10, 2022
Gated MLPs with #Tensorflow and #Keras. medium.com/@radicho/gmlp-… #DeepLearning #AI
gMLP: What it is and how to use it in practice with Tensorflow and Keras?
From medium.com
Radostin Cholakov
@radi_cho
Feb 28, 2023
Next month I will talk about the @TensorFlow TF-DF library and how it can be utilized to combine the strengths of neural networks with decision forest algorithms. In collaboration with @gdg_sofia @gdgbsas @GoogleDevExpert gdg.community.dev/events/details…
476
Radostin Cholakov
@radi_cho
Feb 2, 2023
As an experiment, I am building diffground.com - a simplistic UI to edit images with Stable Diffusion and InstructPix2Pix. It is built with @FlutterDev and is coming to @Android soon (together with more models and editing options).
1K
Radostin Cholakov
@radi_cho
Dec 16, 2022
Today I presented our work titled "Efficient Task-Oriented Dialogue Systems with Response Selection as an Auxiliary Task" at @ICNLSP
904
Radostin Cholakov
@radi_cho
Mar 2, 2023
With the latexify_py package from @Google you can easily generate LaTeX expressions from Python code. github.com/google/latexif…
711
Radostin Cholakov
@radi_cho
Mar 24, 2024
I tried to get assistance with some research work from Gemini 1.5 Pro with a context length of 1 million tokens. Here are some takeaways:
The power of long contexts: Gemini 1.5 Pro use-cases in research.
From medium.com
542
Radostin Cholakov
@radi_cho
Jul 30, 2024
📰 Today, we release ImagiNet, a high-resolution balanced dataset for synthetic image detection. It comes with a strong baseline based on self-contrastive learning, which achieves state-of-the-art results on multiple benchmarks. Some details: - Includes 200K high-resolution
754
Radostin Cholakov
@radi_cho
Dec 22, 2022
We live in a crazy world where prompt marketplaces for pre-trained models are now a thing!
Santiago
@svpino
Dec 22, 2022
You can now buy and sell prompts for generative models: DALL·E, GPT-3, Mid journey, and Stable Diffusion. I'm not sure if "prompt engineering" is a long-term thing, but it's amazing to see how this is a new avenue for creators to make money: promptbase.com
2K
Radostin Cholakov
@radi_cho
Jun 27, 2024
Had a great time with @mervenoyann and many other developers & ML people at yesterday's #GoogleIOConncet community dinner!
320
Radostin Cholakov
@radi_cho
May 10, 2025
lm-eval-harness with the exact same config gives wildly inconsistent results depending on which GPU you use (RTX 4090, A100, H100)??? 🥲 there should really be a better way to benchmark LLMs!
1.2K