Felladrin (Victor Nogueira)

reacted to blaise-tk's post with 🚀 6 days ago

Post

2947

A few months ago, I shared that I was building with @deeivihh something like "the Steam for open source apps"...

🚀 Today, I’m excited to announce that Dione is now open source and live in public beta!

Our mission is simple: make it easier to discover, use, and contribute to open source applications.

🔗 GitHub: https://github.com/dioneapp/dioneapp
💬 Join the community: https://discord.gg/JDFJp33vrM

Want to give it a try? I’d love your feedback! 👀

reacted to fdaudens's post with 👍 2 months ago

Post

1872

Want to know which AI models are least likely to hallucinate — and how to keep yours from spiking hallucinations by 20%?

A new benchmark called Phare, by Giskard, tested leading models across multiple languages, revealing three key findings:

1️⃣ Popular models aren't necessarily factual. Some models ranking highest in user satisfaction benchmarks like LMArena are actually more prone to hallucination.

2️⃣ The way you ask matters - a lot. When users present claims confidently ("My teacher said..."), models are 15% less likely to correct misinformation vs. neutral framing ("I heard...").

3️⃣ Telling models to "be concise" can increase hallucination by up to 20%.

What's also cool is that the full dataset is public - use them to test your own models or dive deeper into the results! H/t @davidberenstein1957 for the link.

- Study: https://www.giskard.ai/knowledge/good-answers-are-not-necessarily-factual-answers-an-analysis-of-hallucination-in-leading-llms
- Leaderboard: https://phare.giskard.ai/
- Dataset: giskardai/phare

reacted to danielhanchen's post with 🚀 2 months ago

Post

5933

🦥 Introducing Unsloth Dynamic v2.0 GGUFs!
Our v2.0 quants set new benchmarks on 5-shot MMLU and KL Divergence, meaning you can now run & fine-tune quantized LLMs while preserving as much accuracy as possible.

Llama 4: unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF
DeepSeek-R1: unsloth/DeepSeek-R1-GGUF-UD
Gemma 3: unsloth/gemma-3-27b-it-GGUF

We made selective layer quantization much smarter. Instead of modifying only a subset of layers, we now dynamically quantize all layers so every layer has a different bit. Now, our dynamic method can be applied to all LLM architectures, not just MoE's.

Blog with Details: https://docs.unsloth.ai/basics/dynamic-v2.0

All our future GGUF uploads will leverage Dynamic 2.0 and our hand curated 300K–1.5M token calibration dataset to improve conversational chat performance.

For accurate benchmarking, we built an evaluation framework to match the reported 5-shot MMLU scores of Llama 4 and Gemma 3. This allowed apples-to-apples comparisons between full-precision vs. Dynamic v2.0, QAT and standard iMatrix quants.

Dynamic v2.0 aims to minimize the performance gap between full-precision models and their quantized counterparts.

reacted to eaddario's post with 🔥 3 months ago

Post

1536

Tensor-wise (TWQ) and Layer-wise quantization (LWQ) now available in llama.cpp!

As of version b5125 users can now do TWQ, whereby you quantize a whole tensor at a specific level, or perform LWQ by choosing specific layers per tensor/s

The new --tensor-type option enables llama-quantize to apply user-defined quant levels to any combination of allowed tensors (i.e. tensors with 2 or more dimensions) and layer number, with support for regex patterns.

For example, to TWQ the Attention Value tensor you would use --tensor-type attn_v=q6_k and to perform LWQ you'll use something like --tensor-type "\.([0-9]|1[01257]|31)\.attn_v=q4_k"

In the next few days/weeks I'll update the models in my HF repo (and will add some others) but eaddario/DeepSeek-R1-Distill-Llama-8B-GGUF and eaddario/DeepSeek-R1-Distill-Qwen-7B-GGUF have been already LWQed.

For reference, compared to the naive Q4_K_M model, the LWQ Qwen-7B is almost 11% smaller (4.68GB vs 4.18GB) with only a 0.35% penalty on PPL!

I'll update the https://medium.com/@eaddario/squeezing-tensor-bits-the-quest-for-smaller-llms-86b23bd052ca post to explain the process in detail, but in the meantime the following links will provide some background:

- Changes to llama-quantize: https://github.com/ggml-org/llama.cpp/pull/12511
- TWQ & LWQ tests: https://github.com/ggml-org/llama.cpp/discussions/12741
- Modified llama-imatrix (not yet merged) used to generate imatrix statistics to guide the TWQ and LWQ process: https://github.com/ggml-org/llama.cpp/pull/12718

reacted to grimjim's post with 🧠 3 months ago

Post

2025

I recently have been looking at a paper titled "Why Warmup the Learning Rate? Underlying Mechanisms and Improvements", by Dayal Singh Kalra and Maissam Barkeshli, and was struck by "warmup" being analogous to simulated annealing.
https://arxiv.org/abs/2406.09405
Taking the physical analogy further, the "warmup" is a stochastic process to knock the system out of current local minima, allowing easier transition toward newer minima. It works because it reduces "fit" and therefore "friction".

reacted to BlinkDL's post with 🔥 3 months ago

Post

8369

RWKV-7 "Goose" 0.4B trained w/ ctx4k automatically extrapolates to ctx32k+, and perfectly solves NIAH ctx16k 🤯 100% RNN and attention-free. Only trained on the Pile. No finetuning. Replicable training runs. tested by our community: https://github.com/Jellyfish042/LongMamba

reacted to mlabonne's post with 🔥 4 months ago

Post

16455

✂️ AutoAbliteration

I made a Colab notebook to automatically abliterate models.

It's quite general, so you can do interesting stuff like blocking a given language in the model outputs.

💻 Colab: https://colab.research.google.com/drive/1RmLv-pCMBBsQGXQIM8yF-OdCNyoylUR1?usp=sharing

1 reply

·

reacted to sharpenb's post with 🔥 4 months ago

Post

3108

We open-sourced the pruna package that can be easily installed with pip install pruna :) It allows to easily ccompress and evaluate AI models including transformers and diffusers.

- Github repo: https://github.com/PrunaAI/pruna
- Documentation: https://docs.pruna.ai/en/stable/index.html

With open-sourcing, people can now inspect and contribute to the open code. Beyond the code, we provide detailed readme, tutorials, benchmarks, and documentation to make transparent compression, evaluation, and saving/loading/serving of AI models.

Happy to share it with you and always interested in collecting your feedback :)

2 replies

·

reacted to AdinaY's post with 🚀 4 months ago

Post

2739

New 3D models from Tencent Hunyuan are now available on the hub 🔥

✨ Hunyuan3D-2mv: multiview shape model for high quality generation
✨ Hunyuan3D-2mini: 0.6B lightweight model for efficient workflows

Model:
tencent/Hunyuan3D-2mv
tencent/Hunyuan3D-2mini
Demo:
tencent/Hunyuan3D-2mv

1 reply

·

reacted to eaddario's post with 👀 4 months ago

Post

2079

Squeezing out tensor bits, part II

At post time, watt-ai/watt-tool-70B continues to top the Berkeley Function-Calling Leaderboard, with the 8B version occupying the 4th place. A remarkable achievement for a model of that size!

The "squeezed" version is now available at eaddario/Watt-Tool-8B-GGUF

(For context please see: https://huggingface.co/posts/eaddario/832567461491467)

2 replies

·

reacted to ngxson's post with 🚀 4 months ago

Post

4376

A comprehensive matrix for which format should you use.

Read more on my blog post: https://huggingface.co/blog/ngxson/common-ai-model-formats

| Hardware        | GGUF      | PyTorch                | Safetensors              | ONNX  |
|-----------------|-----------|------------------------|--------------------------|-------|
| CPU             | ✅ (best) | 🟡                      | 🟡                       | ✅    |
| GPU             | ✅        | ✅                      | ✅                       | ✅    |
| Mobile          | ✅        | 🟡 (via executorch)     | ❌                       | ✅    |
| Apple silicon   | ✅        | 🟡                      | ✅ (via MLX framework)   | ✅    |

1 reply

·

reacted to AdinaY's post with 🔥 4 months ago

Post

2744

Wan2.1 🔥📹 new OPEN video model by Alibaba Wan team!

Model: Wan-AI/Wan2.1-T2V-14B
Demo: Wan-AI/Wan2.1

✨Apache 2.0
✨8.19GB VRAM, runs on most GPUs
✨Multi-Tasking: T2V, I2V, Video Editing, T2I, V2A
✨Text Generation: Supports Chinese & English
✨Powerful Video VAE: Encode/decode 1080P w/ temporal precision

1 reply

·

reacted to JingzeShi's post with 🚀 4 months ago

Post

3016

🤗Welcome to the Doge Edge Device Small language Model.

SmallDoge/Doge-160M-Instruct

reacted to fantos's post with 🚀 5 months ago

Post

6172

😊 Panorama X3 Image

an innovative system that leverages a Stable Diffusion XL-based tiling pipeline to generate unique and vibrant panoramic images by applying different prompts to the left, center, and right sections of a single image.

Key Features & Strengths
Multi-Area Prompt Support
Input distinct descriptions for the left, center, and right regions (e.g., "dense forest" for the left, "calm lake" for the center, and "majestic mountains" for the right). This allows the system to seamlessly blend multiple scenes into one stunning panoramic image. 🌄

Automatic Korean-to-English Translation
If your prompt contains Korean text, it will be automatically translated into English before image generation.
(For example, "안개 낀 산" becomes "Misty mountain") 🔄
This feature ensures that you can effortlessly use both English and Korean prompts.

Advanced Tiling Technology
The project uses a sophisticated tiling approach that manages overlapping regions to produce natural transitions and high-resolution panoramic images.
This isn't just a simple image merge—it's a refined process that delivers exceptional quality and detail. 🖼️

User-Friendly Interface
Enjoy a modern, visually appealing UI featuring a gradient background, semi-transparent containers, and smooth animation effects.
The prompt input fields clearly indicate that both English and Korean entries are allowed with the label (English/Korean allowed), making it accessible for everyone. 🎨

fantos/Panorama

Panorama X3 Image is the perfect tool for anyone looking to visually express creative ideas. Try it out now by experimenting with various prompts and create your very own breathtaking panoramic image! 🚀

Thank you! 🙏

7 replies

·

reacted to Xenova's post with 🚀🔥 5 months ago

Post

13563

We did it. Kokoro TTS (v1.0) can now run 100% locally in your browser w/ WebGPU acceleration. Real-time text-to-speech without a server. ⚡️

Generate 10 seconds of speech in ~1 second for $0.

What will you build? 🔥
webml-community/kokoro-webgpu

The most difficult part was getting the model running in the first place, but the next steps are simple:
✂️ Implement sentence splitting, allowing for streamed responses
🌍 Multilingual support (only phonemization left)

Who wants to help?

11 replies

·

replied to victor's post 5 months ago

This update is massive!! 🙌

I’d love if we could also filter spaces in a way we could list only the ones in Running state.

reacted to Tonic's post with 🔥 5 months ago

Post

3016

🙋🏻‍♂️ Hey there folks ,

our team made a game during the @mistral-game-jam and we're trying to win the community award !

try our game out and drop us a ❤️ like basically to vote for us !

Mistral-AI-Game-Jam/TextToSurvive

hope you like it !