"This section of the plan will also encourage the development of so-called "open-weights" AI models"

🚀 MegaTTS 3 voice cloning is here For context: a while back, ByteDance released MegaTTS 3 (with exceptional voice cloning capabilities), but for various reasons, they decided not to release the WavVAE encoder necessary for voice cloning to work. Recently, a WavVAE encoder compatible with MegaTTS 3 was released by ACoderPassBy on ModelScope with quite promising results. Overall looks quite impressive - excited to see that we can finally do voice cloning with MegaTTS 3! (h/t to MysteryShack on the StyleTTS 2 Discord for info about the WavVAE encoder)

If i could have a wish today i would wish kimi and qwen release their post training datasets like nous does 🫣🤗 We could all be building off eachothers work a lot easier that way!

Has there been no studies on hybrid reasoners? I'm going to be doing some runs across a new model with no non-reasoning dataset with it. Will let you know how reasoning only data impacts the model vs when there is reasoning-less data. Probably a few weeks will have some good insights if I can get the time.

Manifold had much higher odds towards IMO Gold occurring by 2025's end

I memba wen everyone said it was a tokenizer issue

Hermes 3 from @NousResearch is the #1 dataset on @huggingface I plan to use it Don't sleep on Nous They: - Build amazing open source fine-tunes - Built distributed compute - Have well curated datasets - Are fully transparent with model weights and code - Post based takes on X

Its hard to understate how difficult and important good infrastructure is to getting things done

I’ll be presenting TorchTitan: a PyTorch native platform for training foundation models tomorrow at the ICML @ESFoMo workshop! Come and say Hi!

Drones being used to indicate exit for an event in Osaka, Japan