👋 Open to Work

Valeriy Selitskiy PRO

WaveCut

81 158 1418

AI & ML interests

Looking to switch from hobby to career

Recent Activity

liked a model about 2 hours ago

Qwen/Qwen3.8-Flash-Next

liked a model about 2 hours ago

BornSaint/minimax-music3-latent-refiner-v0.10

liked a model about 5 hours ago

ai-sage/Giga-Embeddings-instruct-480M-0826

View all activity

Organizations

reacted to bartowski's post with 👍 4 months ago

Post

59813

You may have noticed that my upload of MiMo-V2.5 upload didn't have the author in the model name:

bartowski/MiMo-V2.5-GGUF

Going forward, I plan to upload models from major 1st party developers without the author name attached for cleanliness, I feel it results in a nicer and more expected user experience

I will continue to uploaded fine tunes with that author + "_" appended for clarity, I personally feel it's nice to know at a glance who's tune it is, but it's also for the reason I first started doing it, to avoid it being confused for a new version of the official release

I hope this change makes sense, it seemed most reasonable to me and a poll I did (forever ago, I move slow sometimes) made it seem likely others would find it reasonable as well (feel free to let me know if you disagree, may not change my mind but I do value knowing what others think)

Thanks for downloading :)

5 replies

reacted to eaddario's post with 👍 7 months ago

Post

3124

Experimental global target bits‑per‑weight quantization of mistralai/Ministral-3-14B-Instruct-2512 and mistralai/Ministral-3-14B-Reasoning-2512

Unlike standard llama.cpp quantizations that rely on fixed type heuristics (e.g., Q4_K_M), the Target BPW approach optimizes per-tensor precision where it matters the most, and produces high quality models that meet a precise global file size target.

Key Advantages:
- VRAM Maximization: Can generate high quality models sized exactly to fit hardware constraints (e.g., fitting the model into exactly 24GB VRAM).
- Data-Driven Precision: Quantization mix is determined by actual weight error sensitivity rather than hardcoded rules, often yielding better PPL/KLD size trade-offs.

Full benchmarks (PPL, KLD, ARC, MMLU, etc.) and methodology in the models' cards

eaddario/Ministral-3-14B-Instruct-2512-GGUF
eaddario/Ministral-3-14B-Reasoning-2512-GGUF

reacted to ovi054's post with 👍 about 1 year ago

Post

3794

WAN 2.2 Text to Image ⚡

ovi054/wan2-2-text-to-image

We all know that WAN 2.2 A14B is a video model. But It turns out this video model can also produce great image results with incredible prompt adherence! The image output is sharp, detailed, and sticks to the prompt better than most.

👉 Try it now: ovi054/wan2-2-text-to-image

1 reply

reacted to cgeorgiaw's post with 🔥 about 1 year ago

Post

1653

Snooping on HF is the best because sometimes you just discover that someone (in this case, Earth Species Project) is about to drop terabytes of sick (high quality animal sounds) data...

EarthSpeciesProject/NatureLM-audio-training

reacted to as-cle-bert's post with 🔥 over 1 year ago

Post

3018

Llama-4 is out and I couldn't resist but to cook something with it... So I came up with 𝐋𝐥𝐚𝐦𝐚𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡𝐞𝐫 (https://llamaresearcher.com), your deep-research AI companion!🔎

The workflow behind 𝗟𝗹𝗮𝗺𝗮𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵𝗲𝗿 is simple:
💬 You submit a query
🛡️ Your query is evaluated by Llama 3 guard model, which deems it safe or unsafe
🧠 If your query is safe, it is routed to the Researcher Agent
⚙️ The Researcher Agent expands the query into three sub-queries, with which to search the web
🌐 The web is searched for each of the sub-queries
📊 The retrieved information is evaluated for relevancy against your original query
✍️ The Researcher Agent produces an essay based on the information it gathered, paying attention to referencing its sources

The agent itself is also built with easy-to-use and intuitive blocks:
🦙 LlamaIndex provides the agentic architecture and the integrations with the language models
⚡Groq makes Llama-4 available with its lightning-fast inference
🔎 Linkup allows the agent to deep-search the web and provides sourced answers
💪 FastAPI does the heavy loading with wrapping everything within an elegant API interface
⏱️ Redis is used for API rate limiting
🎨 Gradio creates a simple but powerful user interface

Special mention also to Lovable, which helped me build the first draft of the landing page for LlamaResearcher!💖

If you're curious and you want to try LlamaResearcher, you can - completely for free and without subscription - for 30 days from now ➡️ https://llamaresearcher.com
And if you're like me, and you like getting your hands in code and build stuff on your own machine, I have good news: this is all open-source, fully reproducible locally and Docker-ready🐋
Just go to the GitHub repo: https://github.com/AstraBert/llama-4-researcher and don't forget to star it, if you find it useful!⭐

As always, have fun and feel free to leave your feedback✨

2 replies

reacted to javelinsoam's post with 🔥 over 1 year ago

Post

2223

Requests Fail with 404 on HuggingFace Inference Due to X-Forwarded-Host Header

We’re encountering a 404 Not Found error from the HuggingFace Inference endpoint when the request includes the X-Forwarded-Host header.

The issue appears to stem from the presence of this header, even if we use any private/public domain:

X-Forwarded-Host: google.com

Without Header – Works

When this header is removed, the request succeeds.
Identical payloads and endpoints return valid responses when the header is omitted.

With Header – Fails

If included (even with a valid public domain), the request fails with:

{
  "error": "Not Found: google.com"
}

You can use curl command to replicate this issue

curl "https://{your-inference-endpoint}/v1/chat/completions" \
-X POST \
-H "Authorization: Bearer <HF_TOKEN>" \
-H "Content-Type: application/json" \
-H "X-Forwarded-Host: any-domain.com" \
-d '{
  "model": "unsloth/DeepSeek-R1-GGUF",
  "messages": [{"role": "user", "content": "What is deep learning?"}],
  "max_tokens": 150
}'

Please let us know if there’s a workaround or config option available to suppress this behavior.

posted an update over 1 year ago

Post

2814

Just posted a new article about YandexGPT 5 family of models

https://huggingface.co/blog/WaveCut/yandexgpt5-models-family-digest

reacted to AdinaY's post with 🔥 over 1 year ago

Post

2950

RWKV7-G1 0.1B 🔥 Pure RNN reasoning model released by RWKV

Model: BlinkDL/rwkv7-g1
paper: RWKV-7 "Goose" with Expressive Dynamic State Evolution (2503.14456)

✨ Apache2.0
✨ Supports 100+ languages
✨ 0.1 B runs smoothly on low power devices
✨ 0.4B/1.5B/2.9B are coming soon!!

1 reply

reacted to sometimesanotion's post with 👍 over 1 year ago

Post

2780

I've managed a #1 score of 41.22% average for 14B parameter models on the Open LLM Leaderboard. As of this writing, sometimesanotion/Lamarck-14B-v0.7 is #8 for all models up to 70B parameters.

It took a custom toolchain around Arcee AI's mergekit to manage the complex merges, gradients, and LoRAs required to make this happen. I really like seeing features of many quality finetunes in one solid generalist model.

8 replies

reacted to grimjim's post with 👍 over 1 year ago

Post

2872

I'm (finally) releasing a Python script that trims excess weights in Gemma2 full-weight models that bloated by ~1B parameters due to an early mergekit bug.
https://github.com/jim-plus/Gemma2-mergekit-remediation

I'd noticed something was off when merges of Gemma2 9B models ended up having ~10B parameters. The current mergekit package is fine, but there are still bloated models on HF that could stand to be fixed.

The script assumes that it will be run from the same directory as the model weights, and will trim the unnecessary lm_head.weight tensor and corresponding index entry.

2 replies

reacted to reach-vb's post with 🔥 almost 2 years ago

Post

5729

Multimodal Ichigo Llama 3.1 - Real Time Voice AI 🔥

> WhisperSpeech X Llama 3.1 8B
> Trained on 50K hours of speech (7 languages)
> Continually trained on 45hrs 10x A1000s
> MLS -> WhisperVQ tokens -> Llama 3.1
> Instruction tuned on 1.89M samples
> 70% speech, 20% transcription, 10% text
> Apache 2.0 licensed ⚡

Architecture:
> WhisperSpeech/ VQ for Semantic Tokens
> Llama 3.1 8B Instruct for Text backbone
> Early fusion (Chameleon)

I'm super bullish on HomeBrew/ Jan and early fusion, audio and text, multimodal models!

(P.S. Play with the demo on Hugging Face: jan-hq/Ichigo-llama3.1-s-instruct)

reacted to nyuuzyou's post with 👀 almost 2 years ago

Post

2022

🎓 Introducing Doc4web.ru Documents Dataset - nyuuzyou/doc4web

Dataset highlights:
- 223,739 documents from doc4web.ru, a document hosting platform for students and teachers
- Primarily in Russian, with some English and potentially other languages
- Each entry includes: URL, title, download link, file path, and content (where available)
- Contains original document files in addition to metadata
- Data reflects a wide range of educational topics and materials
- Licensed under Creative Commons Zero (CC0) for unrestricted use

The dataset can be used for analyzing educational content in Russian, text classification tasks, and information retrieval systems. It's also valuable for examining trends in educational materials and document sharing practices in the Russian-speaking academic community. The inclusion of original files allows for in-depth analysis of various document formats and structures.

reacted to reach-vb's post with 👍 almost 2 years ago

Post

3581

NEW: Open Source Text/ Image to video model is out - MIT licensed - Rivals Gen-3, Pika & Kling 🔥

> Pyramid Flow: Training-efficient Autoregressive Video Generation method
> Utilizes Flow Matching
> Trains on open-source datasets
> Generates high-quality 10-second videos
> Video resolution: 768p
> Frame rate: 24 FPS
> Supports image-to-video generation

> Model checkpoints available on the hub 🤗: rain1011/pyramid-flow-sd3

reacted to tomaarsen's post with 🔥 almost 2 years ago

Post

7302

📣 Sentence Transformers v3.2.0 is out, marking the biggest release for inference in 2 years! 2 new backends for embedding models: ONNX (+ optimization & quantization) and OpenVINO, allowing for speedups up to 2x-3x AND Static Embeddings for 500x speedups at 10-20% accuracy cost.

1️⃣ ONNX Backend: This backend uses the ONNX Runtime to accelerate model inference on both CPU and GPU, reaching up to 1.4x-3x speedup depending on the precision. We also introduce 2 helper methods for optimizing and quantizing models for (much) faster inference.
2️⃣ OpenVINO Backend: This backend uses Intel their OpenVINO instead, outperforming ONNX in some situations on CPU.

Usage is as simple as SentenceTransformer("all-MiniLM-L6-v2", backend="onnx"). Does your model not have an ONNX or OpenVINO file yet? No worries - it'll be autoexported for you. Thank me later 😉

🔒 Another major new feature is Static Embeddings: think word embeddings like GLoVe and word2vec, but modernized. Static Embeddings are bags of token embeddings that are summed together to create text embeddings, allowing for lightning-fast embeddings that don't require any neural networks. They're initialized in one of 2 ways:

1️⃣ via Model2Vec, a new technique for distilling any Sentence Transformer models into static embeddings. Either via a pre-distilled model with from_model2vec or with from_distillation where you do the distillation yourself. It'll only take 5 seconds on GPU & 2 minutes on CPU, no dataset needed.
2️⃣ Random initialization. This requires finetuning, but finetuning is extremely quick (e.g. I trained with 3 million pairs in 7 minutes). My final model was 6.6% worse than bge-base-en-v1.5, but 500x faster on CPU.

Full release notes: https://github.com/UKPLab/sentence-transformers/releases/tag/v3.2.0
Documentation on Speeding up Inference: https://sbert.net/docs/sentence_transformer/usage/efficiency.html

1 reply

reacted to MoritzLaurer's post with ❤️ almost 2 years ago

Post

4790

#phdone - I defended my PhD yesterday! A key lesson: it is amazing how open science and open source can empower beginners with limited resources:

I first learned about instruction-based classifiers like BERT-NLI 3-4 years ago, through the @HuggingFace ZeroShotClassificationPipeline. Digging deeper into this, it was surprisingly easy to find new datasets, newer base models, and reusable fine-tuning scripts on the HF Hub to create my own zeroshot models - although I didn't know much about fine-tuning at the time.

Thanks to the community effect of the Hub, my models were downloaded hundreds of thousands of times after a few months. Seeing my research being useful for people motivated me to improve and upload newer models. Leaving my contact details in the model cards led to academic cooperation and consulting contracts (and eventually my job at HF).

That's the power of open science & open source: learning, sharing, improving, collaborating.

I mean every word in my thesis acknowledgments (screenshot). I'm very grateful to my supervisors @vanatteveldt @CasAndreu @KasperWelbers for their guidance; to @profAndreaRenda and @CEPS_thinktank for enabling me to work part-time during the first year; to @huggingface for creating awesome tools and an awesome platform; and to many others who are not active on social media.

Links to the full thesis and the collection of my most recent models are below.

PS: If someone happens to speak Latin, let me know if my diploma contains some hidden Illuminati code or something :D

4 replies

reacted to ehristoforu's post with ❤️ about 2 years ago

Post

4607

😏 Hello from Project Fluently Team!

✨ Finally we can give you some details about Supple Diffusion. We worked on it for a long time and we have little left, we apologize that we had to increase the work time.

🛠️ Some technical information. The first version will be the Small version (there will also be Medium, Large, Huge, possibly Tiny), it will be based on the SD1 architecture, that is, one text encoder, U-net, VAE. Now about each component, the first is a text encoder, it will be a CLIP model (perhaps not CLIP-L-path14), CLIP was specially retrained by us in order to achieve the universality of the model in understanding completely different styles and to simplify the prompt as much as possible. Next, we did U-net, U-net in a rather complicated way, first we trained different parts (types) of data with different U-nets, then we carried out merging using different methods, then we trained DPO and SPO using methods, and then we looked at the remaining shortcomings and further trained model, details will come later. We left VAE the same as in SD1 architecture.

🙌 Compatibility. Another goal of the Supple model series is full compatibility with Auto1111 and ComfyUI already at the release stage, the model is fully supported by these interfaces and the diffusers library and does not require adaptation, your usual Sampling methods are also compatible, such as DPM++ 2M Karras, DPM++ SDE and others.

🧐 Today, without demo images (there wasn’t much time), final work is underway on the model and we are already preparing to develop the Medium version, the release of the Small version will most likely be in mid-August or earlier.

😻 Feel free to ask your questions in the comments below the post, we will be happy to answer them, have a nice day!

1 reply

reacted to KingNish's post with 👍 over 2 years ago

Post

2969

Something is wrong with GPT-4o

Today, I gained access to GPT-4o, so I thought to test it. However, I encountered several problems, such as When I requested image generation, it did not create any images but only provided links, which are also incorrect. 😥 [Image 1]

Subsequently, I considered that my prompt might be incorrect, I attempted once more with a prompt from OpenAI's examples, but it also did not work. 😥 [Image 2]

Then, I tested its logical reasoning skills, which it failed. I presented a question that an 8b model solved with ease, but GPT-4o could not. 😥 [Image 3]

I also attempted to generate an image from another image, but this too was unsuccessful. [image 4]

Nonetheless, it excels in tasks such as image classification and voice chat.

If you've experienced similar issues, please share them here.