jinaai/jina-clip-v2
Feature Extraction • 0.9B • Updated • 35.6k • 328
Carlo Moro
cnmoro
AI & ML interests
None yet
Recent Activity
liked a model 1 day ago
jinaai/jina-clip-v2 reacted to robtacconelli's post with 🤯 4 days ago
🧬 Midicoth: diffusion-based lossless compression — no neural net, no GPU, no training data
What if reverse diffusion could compress text — without a neural network?
Midicoth brings score-based denoising into classical compression. It treats prior smoothing as forward noise and reverses it with Tweedie's formula on a binary tree — 3 denoising steps, James-Stein shrinkage, applied after all model blending. ~2,000 lines of C, single CPU core.
Beats every dictionary compressor we tested:
enwik8 (100 MB) → 1.753 bpb (−11.9% vs xz, −15% vs Brotli, −24.5% vs bzip2)
alice29.txt → 2.119 bpb (−16.9% vs xz)
Outperforms xz, zstd, Brotli, bzip2, gzip on all inputs
PAQ/CMIX still win with hundreds of models + LSTMs. LLM compressors win with pre-trained knowledge. Midicoth closes the gap with pure statistics — no mixer, no gradient descent, just counting.
The Tweedie denoising layer adds 2.3–2.7% on every file tested — the most consistent component in the ablation. Adding SSE or logistic mixers made things worse. In the online setting, count-based beats gradient-based.
No external dependencies. Fully deterministic. Bit-exact encode/decode. ~60 KB/s throughput.
💻 Code: https://github.com/robtacconelli/midicoth
📄 Paper: https://huggingface.co/papers/2603.08771
⭐ Space: https://huggingface.co/spaces/robtacconelli/midicoth
If you ever wondered whether diffusion ideas belong in data compression — here's proof they do. ⭐ appreciated!Organizations
liked a
model 1 day ago
reactedto fffiloni's post with ❤️ 1 day ago
Post
3645
I brought DALL·E mini back to life 🤖🎨
You can try it here:
fffiloni/dalle-mini-reboot
And I also built a batch version using Hugging Face Jobs (up to 50 images per prompt):
fffiloni/dalle-mini-via-jobs
The goal was to stay close to the original JAX/Flax pipeline, while integrating it with modern tooling (Gradio + Jobs).
It ended up being a fun way to revisit this model — still weird, still fun 😄
You can try it here:
fffiloni/dalle-mini-reboot
And I also built a batch version using Hugging Face Jobs (up to 50 images per prompt):
fffiloni/dalle-mini-via-jobs
The goal was to stay close to the original JAX/Flax pipeline, while integrating it with modern tooling (Gradio + Jobs).
It ended up being a fun way to revisit this model — still weird, still fun 😄
- 2 replies
reactedto robtacconelli's post with 🤯 4 days ago
Post
3655
🧬 Midicoth: diffusion-based lossless compression — no neural net, no GPU, no training data
What if reverse diffusion could compress text — without a neural network?
Midicoth brings score-based denoising into classical compression. It treats prior smoothing as forward noise and reverses it with Tweedie's formula on a binary tree — 3 denoising steps, James-Stein shrinkage, applied after all model blending. ~2,000 lines of C, single CPU core.
Beats every dictionary compressor we tested:
enwik8 (100 MB) → 1.753 bpb (−11.9% vs xz, −15% vs Brotli, −24.5% vs bzip2)
alice29.txt → 2.119 bpb (−16.9% vs xz)
Outperforms xz, zstd, Brotli, bzip2, gzip on all inputs
PAQ/CMIX still win with hundreds of models + LSTMs. LLM compressors win with pre-trained knowledge. Midicoth closes the gap with pure statistics — no mixer, no gradient descent, just counting.
The Tweedie denoising layer adds 2.3–2.7% on every file tested — the most consistent component in the ablation. Adding SSE or logistic mixers made things worse. In the online setting, count-based beats gradient-based.
No external dependencies. Fully deterministic. Bit-exact encode/decode. ~60 KB/s throughput.
💻 Code: https://github.com/robtacconelli/midicoth
📄 Paper: Micro-Diffusion Compression -- Binary Tree Tweedie Denoising for Online Probability Estimation (2603.08771)
⭐ Space: robtacconelli/midicoth
If you ever wondered whether diffusion ideas belong in data compression — here's proof they do. ⭐ appreciated!
What if reverse diffusion could compress text — without a neural network?
Midicoth brings score-based denoising into classical compression. It treats prior smoothing as forward noise and reverses it with Tweedie's formula on a binary tree — 3 denoising steps, James-Stein shrinkage, applied after all model blending. ~2,000 lines of C, single CPU core.
Beats every dictionary compressor we tested:
enwik8 (100 MB) → 1.753 bpb (−11.9% vs xz, −15% vs Brotli, −24.5% vs bzip2)
alice29.txt → 2.119 bpb (−16.9% vs xz)
Outperforms xz, zstd, Brotli, bzip2, gzip on all inputs
PAQ/CMIX still win with hundreds of models + LSTMs. LLM compressors win with pre-trained knowledge. Midicoth closes the gap with pure statistics — no mixer, no gradient descent, just counting.
The Tweedie denoising layer adds 2.3–2.7% on every file tested — the most consistent component in the ablation. Adding SSE or logistic mixers made things worse. In the online setting, count-based beats gradient-based.
No external dependencies. Fully deterministic. Bit-exact encode/decode. ~60 KB/s throughput.
💻 Code: https://github.com/robtacconelli/midicoth
📄 Paper: Micro-Diffusion Compression -- Binary Tree Tweedie Denoising for Online Probability Estimation (2603.08771)
⭐ Space: robtacconelli/midicoth
If you ever wondered whether diffusion ideas belong in data compression — here's proof they do. ⭐ appreciated!
liked a
model 8 days ago
upvoted a collection 8 days ago
liked 2
models 9 days ago
liked a
dataset 9 days ago
updated a
model 9 days ago
liked 4
models 16 days ago
allenai/Olmo-Hybrid-Instruct-DPO-7B
Text Generation • 7B • Updated • 5.4k • 17
Polygl0t/Tucano2-qwen-0.5B-Instruct
Text Generation • 0.5B • Updated • 159 • 1
Polygl0t/Tucano2-qwen-1.5B-Instruct
Text Generation • 2B • Updated • 390 • 1
Polygl0t/Tucano2-qwen-3.7B-Instruct
Text Generation • 4B • Updated • 182 • 1
upvoted a collection 16 days ago
reactedto danielhanchen's post with 🤗 19 days ago
Post
5257
Qwen releases 4 new Qwen3.5 Small models: 0.8B • 2B • 4B • 9B!
Run Qwen3.5-0.8B, 2B and 4B on your phone. Run 9B on 6GB RAM.
The vision reasoning LLMs perform better than models 4x their size.
GGUFs to run: https://huggingface.co/collections/unsloth/qwen35
Guide: https://unsloth.ai/docs/models/qwen3.5
Run Qwen3.5-0.8B, 2B and 4B on your phone. Run 9B on 6GB RAM.
The vision reasoning LLMs perform better than models 4x their size.
GGUFs to run: https://huggingface.co/collections/unsloth/qwen35
Guide: https://unsloth.ai/docs/models/qwen3.5
- 5 replies
liked a
model 19 days ago
upvoted a paper 22 days ago
upvoted an article 23 days ago
Article
Mixture of Experts (MoEs) in Transformers
- +5
•
140
liked a
model 23 days ago
liked a
model 25 days ago
