andito/faster-qwen3-tts-usage
20.5 kB
Andres Marafioti
andito
AI & ML interests
Multimodal models, VLM and TTS
Recent Activity
updated a bucket about 6 hours ago
smolagents/demo-s2s-dashboard new activity about 9 hours ago
smolagents/hf-realtime-voice:Depolyment of Qwen3 tts new activity about 16 hours ago
smolagents/hf-realtime-voice:Pro subscription still showing 10-minute limit in hf-realtime-voiceOrganizations
Posts 9
Post
3298
Finally, our new paper is out! "𝗙𝗶𝗻𝗲𝗩𝗶𝘀𝗶𝗼𝗻: 𝗢𝗽𝗲𝗻 𝗗𝗮𝘁𝗮 𝗜𝘀 𝗔𝗹𝗹 𝗬𝗼𝘂 𝗡𝗲𝗲𝗱"! 🥳
FineVision: Open Data Is All You Need (2510.17269)
If you've ever trained a VLM, you know this problem: nobody shares their data mixtures. It's a black box, making replicating SOTA work impossible.
We wanted to change that.
FineVision unifies 200 sources into 24 million samples. With 17.3 million images and 9.5 billion answer tokens, it's the largest open resource of its kind.
In the paper, we share how we built it:
🔍 finding and cleaning data at scale
🧹 removing excessive duplicates across sources
🤗 decontaminating against 66 public benchmarks
My favorite part is Figure 6 (in the video!). It's our visual diversity analysis. It shows that FineVision isn't just bigger; it's more balanced and conceptually richer than other open datasets.
NVIDIA's Eagle 2 paper highlighted just how critical this visual diversity is, and our results confirm it: models trained on FineVision consistently outperform those trained on any other open dataset on 11 benchmarks!
🎉 To celebrate the paper, I’m also releasing a concatenated and shuffled version of the full dataset! 👉
It’s ready to stream, so you can start training your own models right away:
from datasets import load_dataset
d = load_dataset("HuggingFaceM4/FineVision_full_shuffled", split="train", streaming=True)
print(next(iter(d)))
A big shoutout to the first authors: Luis Wiedmann and Orr Zohar. They are rockstars!
FineVision: Open Data Is All You Need (2510.17269)
If you've ever trained a VLM, you know this problem: nobody shares their data mixtures. It's a black box, making replicating SOTA work impossible.
We wanted to change that.
FineVision unifies 200 sources into 24 million samples. With 17.3 million images and 9.5 billion answer tokens, it's the largest open resource of its kind.
In the paper, we share how we built it:
🔍 finding and cleaning data at scale
🧹 removing excessive duplicates across sources
🤗 decontaminating against 66 public benchmarks
My favorite part is Figure 6 (in the video!). It's our visual diversity analysis. It shows that FineVision isn't just bigger; it's more balanced and conceptually richer than other open datasets.
NVIDIA's Eagle 2 paper highlighted just how critical this visual diversity is, and our results confirm it: models trained on FineVision consistently outperform those trained on any other open dataset on 11 benchmarks!
🎉 To celebrate the paper, I’m also releasing a concatenated and shuffled version of the full dataset! 👉
HuggingFaceM4/FineVision_full_shuffled It’s ready to stream, so you can start training your own models right away:
from datasets import load_dataset
d = load_dataset("HuggingFaceM4/FineVision_full_shuffled", split="train", streaming=True)
print(next(iter(d)))
A big shoutout to the first authors: Luis Wiedmann and Orr Zohar. They are rockstars!
papers about VLM reasoning
-
VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning
Paper • 2504.08837 • Published • 46 -
OpenVLThinker: An Early Exploration to Complex Vision-Language Reasoning via Iterative Self-Improvement
Paper • 2503.17352 • Published • 24 -
Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling
Paper • 2412.05271 • Published • 162 -
R1-Omni: Explainable Omni-Multimodal Emotion Recognition with Reinforcing Learning
Paper • 2503.05379 • Published • 38
papers about VLM reasoning
-
VL-Rethinker: Incentivizing Self-Reflection of Vision-Language Models with Reinforcement Learning
Paper • 2504.08837 • Published • 46 -
OpenVLThinker: An Early Exploration to Complex Vision-Language Reasoning via Iterative Self-Improvement
Paper • 2503.17352 • Published • 24 -
Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling
Paper • 2412.05271 • Published • 162 -
R1-Omni: Explainable Omni-Multimodal Emotion Recognition with Reinforcing Learning
Paper • 2503.05379 • Published • 38
Papers that I've read or want to read
spaces 14
Running
11
HugFut
⚽
Your Hugging Face profile, rated out of 99 as a FUT card
Sleeping
Agents
Ml Intern Someid
🎯
Display an interactive tracking dashboard
Running
Reachy
Reachy Mini Conversation App
🎤
Talk with Reachy Mini!
Sleeping
Agents
Mlintern Convfillq
🎯
Show an interactive tracking dashboard
Running
Reachy
1
Standby
🌙
Ambient standby motion for Reachy Mini.
Running
Featured
93
Parakeet STT Progressive Transcription
🎤
Transcribe speech to text instantly with WebGPU acceleration
models 13
andito/s2s
Updated • 4
andito/smolvlm-encoder-free-finevisionmax
0.1B • Updated • 5
andito/convfill-qwen35-4b-repro
Updated
andito/convfill-qwen35-4b-test
Updated
andito/s2s-endpoint
Updated
andito/nanoVLM
Image-Text-to-Text • 0.2B • Updated • 24 • 2
andito/nanoVLM-500M
Image-Text-to-Text • 0.5B • Updated • 14 • 1
andito/moondream05
Updated
andito/SmolVLM-Base-vqav2
Updated • 5
andito/mlx_summarization
Text Generation • 2B • Updated • 11 • 1
datasets 12
andito/qwentts-cpp-python-wheels
Updated • 1.44k
andito/convfill-qwen35-synthetic
Viewer • Updated • 223 • 40
andito/transformers-pr-triage
Viewer • Updated • 50 • 21 • 1
andito/blog-images
Viewer • Updated • 9 • 437
andito/test_olmocr_documents_qa
Viewer • Updated • 92 • 25
andito/google-landmarks
Viewer • Updated • 1.26M • 965 • 1
andito/Objects365_test
Viewer • Updated • 100 • 25
andito/clevr-math-deduplicated
Viewer • Updated • 70k • 155 • 1
andito/ai2d-merged
Viewer • Updated • 4.86k • 42
andito/mathwriting-google
Viewer • Updated • 656k • 240 • 8