Pretty excited about your lecture! Look forward to it!

I heard u trained a super large model on capybara so I...

Dolphin-2.9-8x22b is in the oven.
fft, deepspeed zero3 param offload, 8k sequence, half the layers are targeted.
This is a significantly improved, filtered dataset. Function calling, agentic, math, dolphin and dolphin-coder.

Google presents Best Practices and Lessons Learned on Synthetic Data for Language Models

Provides an overview of synthetic data research, discussing its applications, challenges, and future directions

arxiv.org/abs/2404.07503

3 bit model for Mac that is quite interesting. Never used 3 bit in MLX before. Definitely worth a try!

Welcome Zephyr 141B to Hugging Chat🔥

🎉A Mixtral-8x22B fine-tune
⚡️Super fast generation with TGI
🤗Fully open source (from the data to the UI)

huggingface.co/chat/models/Hu…

Wow this is new to me! But I have been confident in my model's French but didn't expect it to be somewhat top level.

Wow I really love this sub arena! It provides a more comprehensive eval for sure! Btw what makes me surprised is our model perf in French (my second favorite language) LGTM!🥰

🤔Can we assess agents across various apps & OS w.o. crafting new envs?

OSWorld🖥️: A unified, real computer env for multimodal agents to evaluate open-ended computer tasks with arbitrary apps and interfaces on Ubuntu, Windows, & macOS.

+ annotated 369 real-world computer tasks…

Our improved model in the arena at lmsys and we’ve rolled out to ChatGPT users today — stay tuned for better versions to come

Lgtm! Probably the first finetuned Bixtral

I've been working on integrating E2B to OpenDevin from Junyang Lin and I'm pretty excited where the open source community is heading

The open-source future looks bright

The moat of software AI agents is not the thin wrapper layer (Devin, SWE-Agent), but the underlying LLM. Instead of benchmarking the wrapper, I think SWE-Bench is excellent for evaluating coding LLMs instead:

Hold the agent layer fixed and vary only the LLM backend. Provide all…

Check out our recent paper on Music Pretraining with Transformers. This is a teamwork by lots of awesome collaborators at different institutions.

God bless Justine Tunney's llamacpp kernels,
Mixtral8x22b running CPU ONLY at ~9 tokens per sec.
Yep that's GPT4 class AI.

I'll push out cpu-optimized 4bit/8bit EdgeQuants after benchmarking.

Thanks for the table! Huh this really says something! We'll soon catch up (I hope so)!

Wowowoh u did a great job!

This month is crazy. We just opensourced two models in these two to three weeks but we are already behind now. gotta do something man🥹

Check out our new method for evaluating the quality of generated images, VQAScore! It's simple, runs locally, and is relatively good at evaluation.

IT WORKS! Running Mixtral 8x22B with Transformers! 🔥

Running on a DGX (4x A100 - 80GB) with CPU offloading 🤯