mrfakename
@realmrfakename
Followers
895
Following
82
Media
23
Statuses
502
LLMs, TTS, & Open Source Models on 🤗
California
Joined December 2021
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Pinned Tweet
Excited to launch two new SOTA text-to-speech models on the TTS Arena:
- OpenVoice V2, from
@myshell_ai
- Play․HT 2.0 from
@play_ht
Vote on the TTS Arena to see which one is better!
🧵
0
4
23
Hugging Face announces Cosmo 1B, a fully open sourced Phi competitor with an open sourced dataset. The dataset references various articles and textbooks as "seed data" to generate conversations. Licensed under the Apache 2.0 license.
0
42
242
Outperforming Llama 2 7B and LLaMA 13B with <$100K. 8B sparse MoE model, with 2B activated parameters.
3
20
100
UPDATE: A mod on Mistral Discord server confirmed that the model is not any previous model, it's a completely new model
1
7
97
The TTS Arena now supports four new models:
• VoiceCraft by
@PuyuanPeng
• Parler TTS by
@huggingface
• Vokan
• GPT-SoVITS
Try them out today on the TTS Arena! Vote to see which one is best ⬇️
2
7
50
I'm excited to announce my demo of RWKV Music (v4, MIDI)! The model is Apache 2.0 license, which means it can be used commercially!
0
12
41
Uhhh why can’t Apple just partner with
@huggingface
and make their own 2B model?
EXCLUSIVE: Apple is in talks to build Google’s Gemini AI engine into the iPhone in a potential blockbuster deal.
Here's what we know so far
23
84
262
4
3
32
Over 18,000 votes on the TTS Arena 🤯 ElevenLabs leading but XTTS is close behind! Vote to find your favorite TTS model ⬇️
1
6
28
Parler TTS is a new fully open sourced text-to-speech framework trained by Hugging Face. Describe a speaker you're thinking of, enter text, and listen to the generated audio!
1
3
26
Excited to launch OpenPhonemizer, a BSD-licensed open source phonemizer, based on the amazing DeepPhonemizer!
Many of the leading phonemizers rely on the GPL-licensed espeak.
Achieved satisfactory results when testing on StyleTTS 2. Samples & code ⬇️
1
7
25
Over 22,000 votes on the TTS Arena!
0
4
23
0
0
23
Announcing WizardChatML-7B-v0: a fine-tune of the WizardLM 2 7B model trained on a ChatML dataset.
Try it out:
0
2
22
🧵Today, I'm excited to launch two new models on the TTS Arena: MeloTTS and StyleTTS 2. Both are open sourced, permissively licensed, and highly efficient.
Curious to see how they compare with other leading models? Vote on the TTS Arena ⬇️
0
3
20
We need more open source music generation models. Udio, Suno, and now ElevenLabs are all closed source. Jukebox is older, restrictively licensed, and simply not as good. When will the new open source SOTA arrive?
5
3
20
0
4
20
h2o ai presents danube 2 1.8b, in 3 flavors: base, sft, and chat (dpo).
it outperforms similar models of its size (stablelm/gemma/qwen/phi 1.5) - scores avg 48.75 - and is apache 2.0 licensed.
it's the successor to the original danube 1.8b model.
on hf:
1
3
20
Announcing seewav-gui on Hugging Face Spaces, a GUI/online tool for
@honualx
's amazing seewav library!
Try it out ⬇️
1
5
19
Hugging Face is down… but you can still access the TTS Arena through this direct link!
3
3
17
@yacineMTB
Looks like it might be worse than just retinal damage. Looks like searchers for "blindness" increased significantly as well
0
0
17
Demo of MeloTTS by MyShell AI, creators of OpenVoice. MIT-licensed
2
2
16
Hugging Face repository still performing quite well after over 11,000 commits!
3
2
16
Open access Sora-like model. Designed for human video generation but also works for scenes!
0
4
14
I'm excited to announce the release of my latest model: StyleTTS 2 detector.
This audio classification model aims to detect StyleTTS 2-generated content. Details in 🧵
3
4
14
TTS Arena is trending on Hugging Face! Give it a like to help it reach the top!
1
3
13
Update: it's back!
Here's the official page:
1
3
12
Excited to launch autoregressive models on OpenPhonemizer! The autoregressive models perform slightly better than previous models. Try it out:
2
1
12
MergeKit on Colab is good... but uploads can take a while.
Did you know that you can run Jupyter notebooks directly on a HF CPU Space for ultra-fast uploads (much faster than Colab)?
For example, try running
@maximelabonne
's LazyMergeKit on HF!
0
1
12
2
0
11
This headline is incorrect. Both Phi 1/1.5 and 2 were smaller than Phi 3 🤦
3
0
10
Amphion is super cool - it's like an all-in-one toolkit for audio AI, and permissively licensed!
2
0
10
Just learned a new trick from
@multimodalart
's Gradio space: Minimize the Hugging Face Space header by adding:
header: mini
in the Space README metadata! (
@victormustar
would be nice to have this in the HF Docs)
0
1
10
StyleTTS 2 is outperforming XTTS on the TTS Arena 🤯
It's fast, open source, & permissively licensed!
Try it out ⬇️
0
0
10
@jxmnop
It’s not multiple readmes, it’s just multiple markdown files (ie license, etc), but is a nice improvement
1
0
10
Hugging Face Spaces seems to have been hacked… all my tokens were revoked and I got an email saying they detected unauthorized access - including secrets. They recommend revoking all external API keys as well.
@huggingface
7
1
10
Looking for a 𝐩𝐞𝐫𝐦𝐢𝐬𝐬𝐢𝐯𝐞𝐥𝐲-𝐥𝐢𝐜𝐞𝐧𝐬𝐞𝐝 alternative to 𝚎𝚜𝚙𝚎𝚊𝚔's phonemizer? Try 𝐎𝐩𝐞𝐧𝐏𝐡𝐨𝐧𝐞𝐦𝐢𝐳𝐞𝐫, based on DeepPhonemizer! It's BSD-licensed, fast, and efficient!
1
0
9
@jh3yy
This is super cool! But woudl have to say having this on a page would make it almost impossible to read :)
1
0
9
Releasing the Refusal 🚫 dataset, consisting of instructions from the sourced from the Capybara dataset and synthetically-generated refusals. Train your own Goody 2! Thinking about making a DPO version in the future.
0
1
9
@karpathy
@gui_penedo
You can download 350GB subsets and filter them but the subsets are sorted by year
1
0
9
Just surpassed 10K votes on the TTS Arena! Give it a try:
0
3
9
Discussion regarding Llama 3 license:
0
5
9
Claude prohibits training on outputs
> To develop any products or services that compete with our Services, including to develop or train any artificial intelligence or machine learning algorithms or models.
1
2
8
Cool! 12T tokens, 16x12B MoE, but why so many parameters given the performance?
Introducing DBRX: A New Standard for Open LLM 🔔
💻 DBRX is a 16x 12B MoE LLM trained on 📜 12T tokens
🧠DBRX sets a new standard for open LLMs, outperforming established models on various benchmarks.
Is this thread mostly written by DBRX? Yes!
🧵
22
85
472
2
0
8
OpenAI's models seem quite inefficient (not a criticism just saying...)
GPT-3: 175 billion parameters, outperformed by much smaller open models.
GPT-4: 1.8 *trillion* parameters, almost outperformed by much smaller open models.
What's going on?
5
1
8
0
2
8
Umm… Stable Diffusion 3 isn’t open sourced, is it?
🚨 Narrative Violation
Open-source AI Models Can Be As Good As Proprietary Models.
Stable Diffusion 3 just dropped and easily matches, if not surpasses, DALL-E 3 performance!
Clearly, it's also WAY better than whatever Gemini is using.
Here is the best part - It responds to…
43
58
344
2
0
8
Tencent ARC presents Mistral Pro 8B, similar to their previous Llama Pro 8B. They added 1B parameters and trained it on the Comsopedia dataset from Hugging Face.
1
2
8
@_philschmid
The model is not released under Apache 2.0, only the code. See the bottom of the model card
1
0
8
Run
@maximelabonne
's incredible MergeKit model family tree visualizer on Hugging Face Spaces!
Shocking how complex merges can get!
0
3
7
Chat with 🚫 Refusal using the Hugging Face Inference Widgets! Ask it any question and have it politely refuse. Trained on the publicly available Refusal dataset!
1
2
8
Why is everyone using Google Colab > Kaggle (both owned by Google)?
Kaggle offers 2xT4 or P100 GPUs, or TPUs with enough memory to finetune a 7B model. Additionally, they tell you clearly GPU hour limit.
Colab offers 1xT4 and the limits are quite unclear.
1
2
7
Which one for open source AI?
CC0
9
Apache
17
MIT
35
OpenRAIL
1
6
2
7
Theory (probably wrong) about gpt2-chatbot: since
@sama
edited his tweet from gpt-2 to gpt2 I personally think it might be an improved version of the architecture itself (as in the next generation of gpt)
1
0
7
0
0
7
@ClaustrofobicPi
@erhartford
@Google
"Attempts to override or circumvent safety filters or intentionally drive Gemma or Model Derivatives to act in a manner that contravenes this Gemma Prohibited Use Policy." are banned by the Prohibited Use Policy
4
2
7
A demo of MetaVoice 1B on HF Spaces
Thanks to Hugging Face for the GPU grant!
0
2
7
The BEST 7B open model (according to the Open LLM Leaderboard), Apache 2.0 licensed! Not a chat model, it’s a pre trained model!
After many sleepless night I finally broke the world record for the most powerful 7B model.
The model is yours now, enjoy!
64
176
3K
0
0
6
1
0
6
If only Databricks had done this
0
0
6
You can easily customize it to change the number of bars, the color, & more!
0
0
6
@abacaj
The benchmarks they showed were for function calling and tool usage or something so not surprised if it beats GPT-4 on those specific tasks, but then QLoRA’d Mistrals have been shown to outperform GPT-4 on domain specific tasks as well
1
0
6
If anyone's wondering what happened to TheBloke
0
1
6
So the Sora game engine theory was correct!
0
0
5
🚨 new
@karpathy
repo alert 🚨
llm.c: train gpt models from scratch. currently only supports gpt2, but github repo says llama2/gemma coming soon.
github repo ⬇️
0
3
6
🚀 Just reached 100 likes on the TTS Arena!
1
2
6
Announcing... 🚫 Refusal XL!
I scaled up my 🚫 Refusal dataset (instructions and refusals to perform actions) from ~3K to ~16K interactions!
Check it out on HF:
3
0
6
If you do audio AI and don’t follow
@reach_vb
NGMI!!
1
0
6
MIT licensed!! Also, Microsoft stopped using custom architectures, which means this model can be easily fine tuned with any framework that supports Llama!
Microsoft presents Rho-1: Not All Tokens Are What You Need
RHO-1-1B and 7B achieves SotA results of 40.6% and 51.8% on MATH dataset, respectively — matching DeepSeekMath with only 3% of the pretraining tokens.
repo:
abs:
5
44
311
0
1
6
Pruning of the new Yi 1.5 6B model seems to be working quite well… hopefully it will generate coherent text!
0
0
6
🗒️To all researchers working on Gaussian Splatting: please DO NOT use any source code from the Gaussian Splatting repo, because it will SEVERELY impact the licensing opportunities of your code and prevent it from being truly open sourced (under an OSS license).
3
0
6
New "data consent" feature on
@huggingface
datasets powered by
@spawning_
👀 (doesn't look like it actually removes the images though)
1
0
4
1
0
5