GPT4T retakes the SOTA LLM spot (probably) but also i like this way of announcing things: open source your evals, make it reproduceable and comparable. and **LIGHTWEIGHT**: eleuther’s harness and oai/evals always felt far too heavy of a lift and not flexible. this thing
OPEN SOURCE
-
Open-source program-form implementation of ARC 1 training tasks released
By
–
Just open-sourced by Michael Hodel: a program-form implementation of each of the 400 training tasks of ARC 1, capable of generating more task instances. I expect this will be very useful to everyone working on ARC. GitHub: https://
github.com/michaelhodel/r
e-arc
…
ArXiv: https://
arxiv.org/abs/2404.07353 -
Apple’s 512GB RAM Investment Strategy Strengthens MLX Framework
By
–
If Apple have 512GB of shared RAM/VRAM in the pipeline then their recent investments in the open source MLX framework make even more sense
-
Easy-to-use platform with hosted datasets and pretrained models
By
–
hey you said it! they make getting started really easy 🙂 and their hosted resources are so nice, like datasets & pretrained models
-
HuggingFace Trainer DistributedDataParallel Gradient Bug Fix
By
–
ok so, it was a gradient bug. let me explain with the huggingface trainer, when training on multiple GPUs, sometimes the model is stored as an nn.Module, but sometimes it’s wrapped in this DistributedDataParallel thing I was inadvertently forwarding on the nn.Module in
-
DeepSpeed optimization successfully implemented for AI training
By
–
@winglian wrangled deepspeed till it worked
-
Fine-tuned Mixtral 8x22B Inference Optimization Guide
By
–
Fine-tuned Mixtral 8x22B is ready. Curious what the easiest way to do inference is? Spinning up some GPUs for vLLM right now, is there something better?
-
HashiCorp Accused of Copying MPL-2.0 Licensed Code
By
–
"The code in question can be clearly shown to have been copied from older code under the MPL-2.0 license. HashiCorp seems to have copied the same code itself when they implemented their version of this feature." Pretty convincing summary right there.
-
Hugging Face: Access 80K+ Datasets for Machine Learning
By
–
7. Hugging Face You can find 80K+ datasets on Hugging Face. https://
huggingface.co/datasets