inflatebot (Bot)

reacted to grimjim's post with 👍 18 days ago

Post

3259

Going forward, I will be adopting the term Magnitude-Preserving Orthogonal Ablation (MPOA) for my recent work in mitigating model damage from abliteration. The technique potentially unlocks reasoning capacity previously occupied with safety refusal processing.

For details, start here: https://huggingface.co/blog/grimjim/norm-preserving-biprojected-abliteration

Showcase results: grimjim/gemma-3-12b-it-norm-preserved-biprojected-abliterated (outperforms base instruct on UGI Leaderboard NatInt)

(The existing name, while technically accurate, was a bit of a mouthful.)

2 replies

·

reacted to ZennyKenny's post with 🔥 18 days ago

Post

3141

🎉 Wow. Congratulations @bfirsh and the Replicate team on the CloudFlare acquisition!

✌️ You've really built an incredible ecosystem and product offering and should be super proud.

reacted to grimjim's post with 🔥 20 days ago

Post

5011

Implemented a proof of concept sampler in pure PyTorch and transformers.

Max P consists of a dynamic token filter which applies Winsorization to cap the probabilties of top tokens. Specifically, a base probability in the range of [0,1] is used to cap individual token probability; the sampler then redistributes excess proportionally.

https://github.com/jim-plus/maxp-sampler-poc

Combined with Temperature and Min P, this could represent a more intuitive way of reducing repetition in text generation.

2 replies

·

replied to their post 26 days ago

I would like to make a merge of that size, but unfortunately I haven't found any of the Mistrals to be useful; there are seemingly-intractable problems with the architecture that foil every attempt to make it not repeat the same paragraph over and over, even on OpenRouter where my PC ceases to be a factor. (But like, people like them!! So maybe I have a skill issue, who knows!)

Even if I did, though, it wouldn't be "Mag Mell" exactly. There aren't any similar reagents; especially not since Anthracite evaporated and crestf4ll disappeared.
Of a similar philosophy though, absolutely.

I've been focusing on my physical/mental health this year, and I finally got a family doctor since I posted this so I hope to be active again before too much longer.
Right now I'm experimenting with Gemma3-12B, because for a 12B it's very capable (and I like multimodal models,) and Grimjim's projection-abliteration experiment leads me to believe that I can make it do what I want.

So stay tuned, I suppose. Idk. Primarily I do stuff for me, and I'm trying to get back to that. :P

reacted to DmitryRyumin's post with 🔥 about 1 month ago

Post

3945

🚀💡🌟 New Research Alert - ICCV 2025 (Oral)! 🌟🪄🚀
📄 Title: LoftUp: Learning a Coordinate-based Feature Upsampler for Vision Foundation Models 🔝

📝 Description: LoftUp is a coordinate-based transformer that upscales the low-resolution features of VFMs (e.g. DINOv2 and CLIP) using cross-attention and self-distilled pseudo-ground truth (pseudo-GT) from SAM.

👥 Authors: Haiwen Huang, Anpei Chen, Volodymyr Havrylov, Andreas Geiger, and Dan Zhang

📅 Conference: ICCV, 19 – 23 Oct, 2025 | Honolulu, Hawai'i, USA 🇺🇸

📄 Paper: LoftUp: Learning a Coordinate-Based Feature Upsampler for Vision Foundation Models (2504.14032)

🌐 Github Page: https://andrehuang.github.io/loftup-site
📁 Repository: https://github.com/andrehuang/loftup

🚀 ICCV-2023-25-Papers: https://github.com/DmitryRyumin/ICCV-2023-25-Papers

🚀 Added to the Foundation Models and Representation Learning Section: https://github.com/DmitryRyumin/ICCV-2023-25-Papers/blob/main/sections/2025/main/foundation-models-and-representation-learning.md

📚 More Papers: more cutting-edge research presented at other conferences in the DmitryRyumin/NewEraAI-Papers curated by @DmitryRyumin

🔍 Keywords: #LoftUp #VisionFoundationModels #FeatureUpsampling #Cross-AttentionTransformer #CoordinateBasedLearning #SelfDistillation #PseudoGroundTruth #RepresentationLearning #AI #ICCV2025 #ResearchHighlight

reacted to giadap's post with 👍❤️ 2 months ago

Post

10917

One of the hardest challenges in AI safety is finding the right balance: how do we protect people from harm without undermining their agency? This tension is especially visible in conversational systems, where safeguards can sometimes feel more paternalistic than supportive.

In my latest piece for Hugging Face, I argue that open source and community-driven approaches offer a promising (though not exclusive) way forward.

✨ Transparency can make safety mechanisms into learning opportunities.
✨ Collaboration with diverse communities makes safeguards more relevant across contexts.
✨ Iteration in the open lets protections evolve rather than freeze into rigid, one-size-fits-all rules.

Of course, this isn’t a silver bullet. Top-down safety measures will still be necessary in some cases. But if we only rely on corporate control, we risk building systems that are safe at the expense of trust and autonomy.

Read the blog post here: https://huggingface.co/blog/giadap/preserving-agency

7 replies

·

reacted to ZennyKenny's post with ❤️ 2 months ago

Post

8901

🖤 Probably one of my favorite projects that I've worked on so far, introducing Новояз (Novoyaz).

🛠 One of the first acts of the Bolshevik government after the Russian Revolution was the reform and standardization of the Russian language, which at the time had a non-standard and challenging orthography.

📚 Upon its reform the government launched a nationwide campaign called Ликбез (Likbez), which sought to improve literacy in the country (by the way, it worked, bringing the national literacy rate from <20% in the 1920s to >80% by the 1930s).

‼ While this is a remarkable result that should absolutely be celebrated, it's one that has left behind literally hundreds of thousands if not millions of artifacts using pre-reform Russian orthography.

😓 Researchers and historians are working tirelessly to translate these artifacts to modern Russian so that they may be archived and studied but many have told me that. they are doing this BY HAND (!).

💡 I thought, well this is a perfect use case for OCR and a fine-tuned LLM to step in and help to aid in this important work!

🌏 Introducing НОВОЯЗ (NOVOYAZ)! Powered by ChatDOC/OCRFlux-3B and https://huggingface.co/ZennyKenny/oss-20b-prereform-to-modern-ru-merged, researchers can now convert images of their pre-reform documents to modern Russian orthography using the power of open-source AI!

Check it out and drop a like to support more real-world use cases for open source AI outside of traditional tech-centric domains!

ZennyKenny/Novoyaz

replied to their post 3 months ago

It's basically just the Universal-Light preset with min_p turned up a skosh. This addition helped amend some of the token leakage, which in hindsight, was most likely caused by mixing models with different tokenization bases.

Mistral Nemo is a weird beast, its tokenizer was poorly understood at launch and the Anthracite group released a version of it with its formatting tokens replaced with those from ChatML. Mag Mell combined models based on that (notably Magnum itself from which it gets part of its name) and models based on the original instruction-following model. I may have done something wrong with the MergeKit config, but Toasty and I couldn't hammer it out.

Anyways, we ended up recommending MinP of 0.2 because it keeps the character of the replies while preventing the majority of token problems out to a sufficient context length to fit a whole scene into.

posted an update 3 months ago

Post

835

MN-12B-Mag-Mell inflatebot/MN-12B-Mag-Mell-R1 is a year old next week (Sept. 16)!
TL;DR: AMA in the comments!!

Over the last year, Mag Mell has enjoyed a consistent presence in the roleplay scene, with 100,000 downloads total on the full-precision weights, and likely more than that (although I don't control those repos so I can't see) on the quantized versions. (Not bad for a craft-produced merge spun up by someone who barely understood what a tensor was at the time...)

Since then, I've become a de-facto community manager for the

allura-org organization, a group who I'm immensely proud of for what we've achieved in the last year - a space built by and for queer, trans and gender-nonconforming members of the AI space across different use cases. (This has taken up basically all my mental bandwidth, which is the primary reason I haven't done more LLMs! That and Factorio. IYKYK.)

When I picked my name, I expected to burn it in a couple months' time, but instead it's become something of a rent-lowering bulwark, attracting the type of people we want and deflecting the ones we don't, and I've felt more comfortable being my whole weirdo self in the space as a result. While this hasn't resulted in much that I can share on Hugging Face, it's regardless been a very fun and fulfilling time.

If anybody reading is a fan of MM or anything else that I do or just have something funny to ask, feel free to leave any questions in the comments. I may reply directly, or I may collect them into a blog post on the 16th. We'll see how the next week goes for me.

Thank you so much for everything!

9 replies

·

reacted to MonsterMMORPG's post with 🚀 3 months ago

Post

3773

Qwen Image LoRA trainings Stage 1 results and pre-made configs published - As low as training with 6 GB GPUs - Stage 2 research will hopefully improve quality even more - Images generated with 8-steps lightning LoRA + SECourses Musubi Tuner trained LoRA in 8 steps + 2x Latent Upscale

1-click to install SECourses Musubi Tuner app and pre-made training configs shared here : https://www.patreon.com/posts/137551634

Hopefully a full video tutorial will be made after Stage 2 R&D trainings completed

Example training made on the hardest training which is training a person and it works really good. Therefore, it shall work even much better on style training, item training, product training, character training and such

Stage 1 took more than 35 unique R&D Qwen LoRA training

1-Click installer currently fully supporting Windows, RunPod (Linux & Cloud) and Massed Compute (Linux & recommend Cloud) training for literally every GPU like RTX 3000, 4000, 5000 series or H100, B200, L40, etc

28 images weak dataset is used for this training

More angles having dataset would perform definitely better

Moreover, i will make a research for a better activation token as well rather than ohwx

After Stage 2, I am expecting hopefully much better results

As a caption, i recommend to use only ohwx nothing else, not even class token

Higher quality more images shared here : https://medium.com/@furkangozukara/qwen-image-lora-trainings-stage-1-results-and-pre-made-configs-published-as-low-as-training-with-ba0d41d76a05

Image prompts randomly generated with Gemini 2.5 in Google AI Studio for free

How to Generate Images
In the zip file of this post : https://www.patreon.com/posts/114517862

We have Amazing_SwarmUI_Presets_v21.json made for SwarmUI

Import it and i am using Qwen Image 8 Steps Ultra Fast to generate images and then apply Upscale Images 2X to make them 4x resolution (1328x1328 to 2656x2656)

Of course in addition to preset don't forget to select your trained LoRA

1 reply

·

replied to ProCreations's post 4 months ago

LMAO
They spared me, I guess xD

replied to ProCreations's post 4 months ago

I think a bunch of bot accounts got nuked or something.

replied to appvoid's post 4 months ago

If AGI ever does happen, it will be open-source, eventually. There is no One True AGI. Literally some other lab could just make their own AGI afterwards especially because they'll have AGI to help them do it.

It's a weird question to ask. Like if AGI is inevitable, open AGI is also inevitable.

replied to ovi054's post 4 months ago

This, being charitable, is just prompt-oriented image captioning. It doesn't "extract" anything from anything. It provides decent descriptions, but only as good as Florence 2 can do.
There's nothing wrong with any of that inherently, it can still be useful, but claiming it to be a "prompt extractor" is misleading IMO.

Also most of these reactions are from... yourself?

reacted to giadap's post with ❤️ 4 months ago

Post

3300

💬 From Replika to everyday chatbots, millions of people are forming emotional bonds with AI, sometimes seeking comfort, sometimes seeking intimacy. But what happens when an AI tells you "I understand how you feel" and you actually believe it?

At Hugging Face, together with @frimelle and @yjernite , we dug into something we felt wasn't getting enough attention: the need to evaluate AI companionship behaviors. These are the subtle ways AI systems validate us, engage with us, and sometimes manipulate our emotional lives.

Here's what we found:
👉 Existing benchmarks (accuracy, helpfulness, safety) completely miss this emotional dimension.
👉 We mapped how leading AI systems actually respond to vulnerable prompts. 👉 We built the Interactions and Machine Attachment Benchmark (INTIMA): a first attempt at evaluating how models handle emotional dependency, boundaries, and attachment (with a full paper coming soon).

Check out the blog post: https://huggingface.co/blog/giadap/evaluating-companionship

🚢 We also shipped two visualization tools with Gradio to see how different models behave when things get emotionally intense:
- AI-companionship/intima-responses-2D
- giadap/INTIMA-responses

posted an update 5 months ago

Post

1354

Pressure from far-right activist groups is inducing payment processors to crack down on adult content on mainstream platforms (where previously they were restricted in scope to adult platforms), starting first with Steam, now recently itch.io. We saw similar on Civit recently, except they were unlucky enough to have partnered with a payment processor who didn't care that they complied and pulled out anyway. A friend of mine recently lost all her income because of similar stuff (and the Itch move badly hurt other people in my circles.)

This is pretty far-removed from a lot of the people reading this, andI don't want to tell you how to feel about this happening, *but all the same, don't assume that Hugging Face is going to be safe because it's an "academic-coded" platform.* If you make things for mature audiences here, please make sure you have local backups of your data and models.

That's all I got. Be safe out there.

reacted to giadap's post with 👍🤗 5 months ago

Post

1269

🤖 Technology means power, and whoever owns the technology owns the power.

Thrilled to share insights from my recent interview with MIT Technology Review about the growing movement toward local LLMs and what it means for AI democratization. Read here: https://www.technologyreview.com/2025/07/17/1120391/how-to-run-an-llm-on-your-laptop/

🤔 Why this matters: When we use "free" online AI services, we're often the product. Our conversations become training data, our personal stories get "cooked into" models, and our privacy becomes a commodity. But there's an alternative path forward.

💡 The power shift is real: Local LLMs aren't just about privacy; they're about redistributing AI power away from a handful of tech giants. When individuals, organizations, and even entire nations can run their own models, we're democratizing access to AI capabilities.

🤗 At Hugging Face, we're proud to be at the center of this transformation. Our platform hosts the world's largest library of freely downloadable models, making cutting-edge AI accessible to everyone -- from researchers and developers to curious individuals who want to experiment on their laptops or even smartphones.

The technical barriers that once required $$$ server racks are crumbling. Today, anyone with basic computer skills can download a model, run it locally, and maintain complete control over their AI interactions. No sudden algorithm changes, no data harvesting, no corporate gatekeeping.

This is about technical convenience, but especially about technological sovereignty. When AI power is concentrated in a few hands, we risk creating new forms of digital dependency. Local models offer a path toward genuine AI literacy and independence.

🚀 The future of AI should be open, accessible, and in the hands of the many, not the few. What are your thoughts on AI democratization? Have you experimented with local models yet?

Bot PRO

AI & ML interests

Recent Activity

Organizations

Bot PRO

AI & ML interests

Recent Activity

Organizations

inflatebot's activity