From 28250d988474db962dc8b2aef07211f52eeb2b4b Mon Sep 17 00:00:00 2001 From: eckartal <159995642+eckartal@users.noreply.github.com> Date: Thu, 8 Aug 2024 10:17:39 +0300 Subject: [PATCH] Update can-llama-3-listen.mdx (#58) Updated the Hugging Face links. --- src/pages/blog/can-llama-3-listen.mdx | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/src/pages/blog/can-llama-3-listen.mdx b/src/pages/blog/can-llama-3-listen.mdx index 4e75edc..a3eef0f 100644 --- a/src/pages/blog/can-llama-3-listen.mdx +++ b/src/pages/blog/can-llama-3-listen.mdx @@ -20,7 +20,7 @@ import ResearchCTABlog from '@/components/Blog/ResearchCTABlog' -We invite you to join [llama3-s](https://discord.gg/Q7dXd6b9Zb): an open and ongoing research experiment focused on teaching llama3 to listen. This blog post is a quick project update and an open call to join to us to build a speech adapter for open source models. +We invite you to join [llama3-s](https://discord.gg/Q7dXd6b9Zb): an open and ongoing research experiment focused on teaching llama3 to listen. This blog post is a quick project update and an open call to join us in building a speech adapter for open-source models. Inspired by the [Chameleon](https://arxiv.org/pdf/2405.09818) and [Llama Herd](https://arxiv.org/pdf/2407.21783) papers, llama3-s is an early-fusion, audio and text, multimodal model. We're conducting this research entirely in the open, with an open-source [codebase](https://github.com/homebrewltd/llama3-s), [open data](https://huggingface.co/datasets/homebrewltd/instruction-speech-v1.5) and [open weights](https://huggingface.co/homebrewltd/llama3-s-2024-07-19). @@ -80,9 +80,9 @@ You can find the datasets here: | Date | HF Checkpoint | Tokens | | --- | --- | --- | -| 📅 2024-07-19 | 🔗 https://huggingface.co/homebrew-research | 🔢 1.35B | -| 📅 2024-07-18 | 🔗 https://huggingface.co/datasets/homebrew-research/instruction-speech-v1.5 | 🔢 800M | -| 📅 2024-06-30 | 🔗 https://huggingface.co/datasets/homebrew-research/instruction-speech-v1 | 🔢 450M | +| 📅 2024-07-19 | 🔗 https://huggingface.co/homebrewltd | 🔢 1.35B | +| 📅 2024-07-18 | 🔗 https://huggingface.co/datasets/homebrewltd/instruction-speech-v1.55 | 🔢 800M | +| 📅 2024-06-30 | 🔗 https://huggingface.co/datasets/homebrewltd/instruction-speech-v1 | 🔢 450M | **Training**: The instruct tuning was done with fsdp2 ([Torchtune](https://github.com/pytorch/torchtune)) mixed-precision, on a [llama3 8b](https://huggingface.co/meta-llama/Meta-Llama-3-8B) base model, with the final weights in bf16. We used the [AdamMini](https://arxiv.org/abs/2406.16793) optimizer, a global batchsize of 128 (mini-batches of 2-4), a 3e-4 learning rate, and a slightly longer warm up ratio. You can find the full steps to reproduce our training here on [Hugging Face](https://huggingface.co/homebrewltd/llama3-s-2024-07-19).