Hugging Face's logo

Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up

All HF Hub posts
sergiopaniego 
posted an update 1 day ago
view post
Post
2,750
TRL now supports MOPD (Multi-Teacher On-Policy Distillation) natively❗️

Many frontier open models use it. labs train one RL expert per domain and then distill them all into one model

Blog with what it is, who uses it and an experiment you can run on HF Jobs with one command

https://huggingface.co/blog/sergiopaniego/mopd-trl
SeaWolf-AI 
posted an update 3 days ago
view post
Post
5,680
🛑 What if the kill switch lived inside the model itself?
For 50 years, the operating system lived inside the computer. Now the operating system lives inside the AI.
Today we're sharing ViKernel : Weight-Native OS, a Secure OS wired directly into the weights of an LLM.
🔧 How it works
ViKernel wires CPU, NPU (vector unit), RAM and ROM circuits into the model weights, runs them accelerated on GPU or NPU, and boots a Secure OS on top.
⚡ Boots in the weights: change a single byte and it locks itself
⚖️ Decides in the weights: every action is allowed or blocked by rule, and persuasion does not change the verdict
🔗 Remembers in the weights: every decision is sealed in a tamper-evident chain
🛑 Stops in the weights: anyone can stop it, and restarting requires a signed key
The original weights stay untouched, so the model keeps its skills. It plugs into any LLM or VLM.
Why the kill switch belongs inside the model: external guardrails can be removed, argued with, or rewritten. A switch that travels with the weights cannot be left behind. We see it as one layer of defense in depth, alongside hardware stops where you have them.
📖 Read the full write-up:
https://huggingface.co/blog/FINAL-Bench/vikernel-why-the-kill-switch-must-live-inside-the
🖥️ Try the desktop demo (Terminal, Kill Switch Lab, Leak Test, Audit Viewer, and a Kernel Guard game):
FINAL-Bench/ViKernel-Desktop
Try it in 30 seconds: press 🛑 KILL, type run weather (refused), then resume --key vikernel-demo.
tardellirs 
posted an update 1 day ago
view post
Post
2,225
Model cards say what a model is: a fine-tune of X, a quantization of Y. I built Weightprint to check that against the weights.

App: tardellirs/weightprint
Data (every pair, and the copy groups): tardellirs/weightprint

It reads the header and a few bytes of every tensor (no downloads) and compares a model with the parent its card declares. 120,976 models so far. Some of what came out:

- 12% of declared fine-tunes hold their parent's weights unchanged.
- 23% of declared fine-tunes are LoRA-style: only the attention and MLP matrices moved, norms and embeddings untouched. That share went from 10% in 2023 to 25% in 2026.
- Of 1,668 fine-tunes named abliterated, uncensored or heretic, 53% edit only the matrices that write to the residual stream, the signature of an abliteration. The rest were trained further, merged, or are the parent's weights.
- 51% of the weight bytes on the Hub are byte-for-byte repeats of another repo's files.

Search any Hub id to see which layers and parts changed against its parent, or compare any two models.

A sample per tensor can miss things, and some formats can't be compared yet. If a verdict on a model you know looks wrong, tell me here and I'll check it against the engine. That's what this post is for, before a longer write-up.
  • 2 replies
·
savoji 
posted an update 2 days ago
view post
Post
4,728
LFM3-1.7B from @LiquidAI is the #1 trending model on the Hub (855 likes). It takes about 1 GB at 4-bit, and in the vendor's own table it beats Qwen3.5 9B on several agent benchmarks.

The edge decision moves from "are small models good enough?" to "which architecture and runtime do I pick?" One 1.7B backbone now ships as text, VL-2B and Omni-2B checkpoints with open weights.

🔍 Architecture and speed
• 24 layers: 13 gated convolutions (kernel 7), 5 sliding-window attention (256 tokens), 6 global attention, chosen by a search over 14 device and backend targets
• llama.cpp Q4_0 at 32K context: 1.8 to 2.8x faster decode than the next model on 3 devices
• Full 128K context needs 1.6 GB of KV cache vs 4.3 GB (Qwen3.5-4B) and 5.6 GB (MiniCPM5-2B)

📊 Liquid AI's table
• Average 49.37 vs 47.24 for Qwen3.5 9B; BFCL V4 71.39 vs 62.83; WebArena-Verified 38.42 vs 28.94
• Qwen3.5 9B still wins SWE-bench Verified (51.20 vs 43.40) and GPQA Diamond (80.40 vs 63.37)

🛠️ Try it
• GGUF (174,177 downloads last month) runs in llama.cpp b11960+, Ollama and LM Studio; MLX 4-bit for Apple Silicon; SGLang for GPUs
• llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M --temp 0.1 --top-k 50 --repeat-penalty 1.1
• Omni-2B runs in the browser in the WebGPU Space

⚠️ Caveats
• The numbers are vendor-reported, with no independent reproduction in the sources. The card says it is not for agentic coding or complex multi-step reasoning.

LiquidAI/LFM3-1.7B
LiquidAI/LFM3-1.7B-GGUF
LiquidAI/LFM3-Omni-2B
LiquidAI/LFM3-Omni-2B-WebGPU

Has a sub-2B model held up on your own tool-calling harness?

• Medium: https://medium.com/p/05507e4e81c8
• Substack: https://sajjadpakdamansavoji.substack.com/p/liquid-ais-lfm3-turns-the-small-model
• X: https://x.com/saavoji/status/2117307514177941672
• Website: https://sajjadpsavoji.github.io/blog#liquid-ais-lfm3
  • 12 replies
·
gorrce 
posted an update 2 days ago
view post
Post
4,950
Introducing NeonLib — an Android knowledge library with on-device document retrieval powered by Google's EmbeddingGemma 300M and LiteRT.

Import notes and documents, recognise text from photos or scanned PDFs, and search for relevant source passages directly on your phone.

EmbeddingGemma handles retrieval. Optional answers use a separate downloaded local language model or a selected online provider.

Built with Kotlin, Jetpack Compose and ObjectBox vector search. I'm particularly interested in Croatian and English retrieval and making on-device AI useful in everyday document workflows.

Try NeonLib:
https://play.google.com/store/apps/details?id=com.neonbreak.neonlib&hl=en

Feedback from the Hugging Face community is welcome!
  • 4 replies
·
iliya-radulov 
posted an update 1 day ago
view post
Post
2,354
Field Notes from a Low-Cost Robot Learning Lab: Hardware Churn, Software Whiplash, and What Actually Went Wrong

An informal companion piece documenting the hardware incidents, framework churn, and debugging dead-ends from roughly seven weeks of work. It is written for people building similar setups on a hobbyist budget and is candid about the time lost to causes like loose power cables and swapped ports.

Zenodo: https://zenodo.org/records/23259698

About me: I am an independent researcher. These works are part of a single-operator project on low-cost robot learning. My ORCID is 0000-0001-8943-5083. For questions or collaboration, feel free to reach out via the HF community.
mahwizzzz 
posted an update 3 days ago
kutomanovhennadii 
posted an update 1 day ago
view post
Post
2,325
How much load can a language model handle?

“Summarize a Wikipedia article in at least 300 words. Use no commas. Split your answer into three titled sections.”

This is the kind of task used in IFEval, a benchmark of how well models follow multiple requirements at once.

Suppose the model breaks one requirement. Did it lose track of an instruction? Did the no-comma rule make the summary harder to write? Or was summarizing the article difficult to begin with? A program can tell us what failed, but not separate instruction load from task difficulty and interacting constraints.

There is another approach: synthetic tests where a model retrieves a key's value from a long list. Those are easier to control, but they test information retrieval rather than the execution of multiple instructions.

So how can we measure the load itself?

That's the question behind my new article on the IEC instrument. Instead of an essay, the model gets meaningless text made of random words. Instead of complex constraints, it gets simple replacement rules such as lake -> stone. Its job is to rewrite the text while applying every rule.

The simplicity serves a purpose. Each rule applies to its own word, so we can add instructions without changing the operation. We can also set how many different words the text has and how often each repeats. And every application has a known correct outcome.

A single response thus produces many individually verifiable trials, not just one overall score. This lets us study how a model's behavior degrades under different kinds of load.

The article walks through the instrument, an actual prompt and GPT-4o mini response, and how to reproduce the trials. Code and data are public.

Article: https://huggingface.co/blog/kutomanovhennadii/how-to-measure-a-language-models-behavior-under-lo
Data: kutomanovhennadii/iec-gpt-4o-mini-2024-07-18

Originally written in Russian. English translation assisted by AI.
  • 1 reply
·
Banaxi-Tech 
posted an update 1 day ago
tardellirs 
posted an update 3 days ago
view post
Post
5,512
I really like the LocalLLaMA/typed-decisions benchmark, so I built a small Space to make its leaderboard easier to explore: tardellirs/typed-decisions-explorer

It reads the dataset card on every visit, so new submissions show up as soon as they are merged. It has a sortable ranking and two Pareto charts: accuracy × size and accuracy × latency.

A few things it makes easy to see in the zero-shot table today:
- Only three models are within noise of the top: Quyet-1.0-Large (0.791), meraGPT Decider 1 and Baseten Mammoth 1 (0.768).
- The smallest model above Jev 1.13 (0.727) is d3-mini, at 4B.
- The fastest local run above Jev is FastBrain-9B: 50 ms on an RTX PRO 6000.
- 15 of 41 models score above the teacher's self-agreement (0.735), so the top of the table measures agreement with the teacher more than correctness.

Congrats to @codelion and @sandeep85 , the maintainers, for a great benchmark. Feedback welcome!
  • 11 replies
·
Recently active users