sentence-transformers/all-MiniLM-L6-v2
Sentence Similarity • 22.7M • Updated • 197M • • 4.65k
Roberto Tacconelli PRO
robtacconelli
AI & ML interests
None yet
Recent Activity
liked a model 1 day ago
sentence-transformers/all-MiniLM-L6-v2 authored a paper 13 days ago
Aperiodic Structures Never Collapse: Fibonacci Hierarchies for Lossless CompressionOrganizations
None yet
liked a
model 1 day ago
authored a
paper 13 days ago
upvoted 2
papers 13 days ago
submitted a
paper to Daily Papers 13 days ago
liked a
Space 20 days ago
reactedto their post with 🤯🚀 21 days ago
Post
3702
🧬 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!
posted an update 21 days ago
Post
3702
🧬 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!
updated a
Space 22 days ago
published a
Space 22 days ago
liked a
model 23 days ago
authored a
paper 25 days ago
submitted a
paper to Daily Papers 26 days ago
Improve dataset card: add metadata, links and usage information
1
#1 opened about 1 month ago
by 
nielsr
updated a
dataset about 1 month ago
reactedto their post with 🔥 about 1 month ago
Post
3640
🏆 Nacrith: a 135M model that out-compresses everything on natural language
What if a tiny LM could compress english text better than _every_ compressor out there — classical or neural, small or large?
Nacrith pairs SmolLM2-135M with an ensemble of online predictors and high-precision arithmetic coding.
What's inside
The standard LLM+arithmetic coding approach wastes ~75% of CDF precision on large vocabularies. Our CDF-24 fix alone recovers 0.5 bpb. On top: a token N-gram that skips the GPU on predictable tokens, an adaptive bias head, llama.cpp backend (7× faster than PyTorch), multi-GPU parallel compression, and a binary file format (NC06) — the first LLM-based binary compressor we know of.
Runs on a GTX 1050 Ti. ~500 MB weights, ~1.2 GB VRAM per worker.
💻 Code: https://github.com/robtacconelli/Nacrith-GPU
⭐ Space: robtacconelli/Nacrith-GPU
📄 Paper: Nacrith: Neural Lossless Compression via Ensemble Context Modeling and High-Precision CDF Coding (2602.19626)
Try it, break it, share your results — all feedback welcome. ⭐ on the repo appreciated!
Results across all systems we tested:
- alice29.txt → 0.918 bpb (−44% vs CMIX, −20% vs ts_zip) — below the 2nd-order Shannon entropy bound
- enwik8 (100 MB) → 0.9389 bpb (−8% vs FineZip/LLMZip's 8B model, −15% vs ts_zip)
- Unseen text → 0.723 bpb on a doc published after training cutoff — no memorization, 26% better than FineZip/LLMZip on the same model
SmolLM2-135M by
HuggingFaceTB
What if a tiny LM could compress english text better than _every_ compressor out there — classical or neural, small or large?
Nacrith pairs SmolLM2-135M with an ensemble of online predictors and high-precision arithmetic coding.
What's inside
The standard LLM+arithmetic coding approach wastes ~75% of CDF precision on large vocabularies. Our CDF-24 fix alone recovers 0.5 bpb. On top: a token N-gram that skips the GPU on predictable tokens, an adaptive bias head, llama.cpp backend (7× faster than PyTorch), multi-GPU parallel compression, and a binary file format (NC06) — the first LLM-based binary compressor we know of.
Runs on a GTX 1050 Ti. ~500 MB weights, ~1.2 GB VRAM per worker.
💻 Code: https://github.com/robtacconelli/Nacrith-GPU
⭐ Space: robtacconelli/Nacrith-GPU
📄 Paper: Nacrith: Neural Lossless Compression via Ensemble Context Modeling and High-Precision CDF Coding (2602.19626)
Try it, break it, share your results — all feedback welcome. ⭐ on the repo appreciated!
Results across all systems we tested:
- alice29.txt → 0.918 bpb (−44% vs CMIX, −20% vs ts_zip) — below the 2nd-order Shannon entropy bound
- enwik8 (100 MB) → 0.9389 bpb (−8% vs FineZip/LLMZip's 8B model, −15% vs ts_zip)
- Unseen text → 0.723 bpb on a doc published after training cutoff — no memorization, 26% better than FineZip/LLMZip on the same model
SmolLM2-135M by
- 1 reply
upvoted an article about 1 month ago
Article
Nacrith: How a 135M Language Model Became the Best Text Compressor We've Tested
•
1
published an article about 1 month ago
Article
Nacrith: How a 135M Language Model Became the Best Text Compressor We've Tested
•
1
reactedto their post with 🚀 about 1 month ago
Post
3640
🏆 Nacrith: a 135M model that out-compresses everything on natural language
What if a tiny LM could compress english text better than _every_ compressor out there — classical or neural, small or large?
Nacrith pairs SmolLM2-135M with an ensemble of online predictors and high-precision arithmetic coding.
What's inside
The standard LLM+arithmetic coding approach wastes ~75% of CDF precision on large vocabularies. Our CDF-24 fix alone recovers 0.5 bpb. On top: a token N-gram that skips the GPU on predictable tokens, an adaptive bias head, llama.cpp backend (7× faster than PyTorch), multi-GPU parallel compression, and a binary file format (NC06) — the first LLM-based binary compressor we know of.
Runs on a GTX 1050 Ti. ~500 MB weights, ~1.2 GB VRAM per worker.
💻 Code: https://github.com/robtacconelli/Nacrith-GPU
⭐ Space: robtacconelli/Nacrith-GPU
📄 Paper: Nacrith: Neural Lossless Compression via Ensemble Context Modeling and High-Precision CDF Coding (2602.19626)
Try it, break it, share your results — all feedback welcome. ⭐ on the repo appreciated!
Results across all systems we tested:
- alice29.txt → 0.918 bpb (−44% vs CMIX, −20% vs ts_zip) — below the 2nd-order Shannon entropy bound
- enwik8 (100 MB) → 0.9389 bpb (−8% vs FineZip/LLMZip's 8B model, −15% vs ts_zip)
- Unseen text → 0.723 bpb on a doc published after training cutoff — no memorization, 26% better than FineZip/LLMZip on the same model
SmolLM2-135M by HuggingFaceTB
What if a tiny LM could compress english text better than _every_ compressor out there — classical or neural, small or large?
Nacrith pairs SmolLM2-135M with an ensemble of online predictors and high-precision arithmetic coding.
What's inside
The standard LLM+arithmetic coding approach wastes ~75% of CDF precision on large vocabularies. Our CDF-24 fix alone recovers 0.5 bpb. On top: a token N-gram that skips the GPU on predictable tokens, an adaptive bias head, llama.cpp backend (7× faster than PyTorch), multi-GPU parallel compression, and a binary file format (NC06) — the first LLM-based binary compressor we know of.
Runs on a GTX 1050 Ti. ~500 MB weights, ~1.2 GB VRAM per worker.
💻 Code: https://github.com/robtacconelli/Nacrith-GPU
⭐ Space: robtacconelli/Nacrith-GPU
📄 Paper: Nacrith: Neural Lossless Compression via Ensemble Context Modeling and High-Precision CDF Coding (2602.19626)
Try it, break it, share your results — all feedback welcome. ⭐ on the repo appreciated!
Results across all systems we tested:
- alice29.txt → 0.918 bpb (−44% vs CMIX, −20% vs ts_zip) — below the 2nd-order Shannon entropy bound
- enwik8 (100 MB) → 0.9389 bpb (−8% vs FineZip/LLMZip's 8B model, −15% vs ts_zip)
- Unseen text → 0.723 bpb on a doc published after training cutoff — no memorization, 26% better than FineZip/LLMZip on the same model
SmolLM2-135M by
- 1 reply