Correctness is a gradient and I want to backprop it
MACHINE LEARNING
-
Writing Stream Processors: Avoiding the Duck Syndrome
By
–
So You Want to Write a Stream Processor? Beware of the Duck Syndrome. https://
bit.ly/4bCdgTl
#AI #MachineLearning #DeepLearning #LLMs #DataScience -
MMLU Evaluation Method: Log Probability vs Multiple Choice
By
–
Nice! Btw it's possible (in principle) to also evaluate MMLU in the same way I evaluate HellaSwag, where you swap out the 4 continuations in turn and predict the one with highest average log prob. Though it hurts the model by a few percent because it can't reason by elimination.
-
Tetris Computational Complexity: Understanding NP-Hard Problem
By
–
Tetris is NP-Hard – http://
koaning.io https://
bit.ly/3XSJNkQ
#AI #MachineLearning #DeepLearning #LLMs #DataScience -
BAIR Researchers Win Test of Time Award at ICML 2024
By
–
Congratulations to BAIR faculty and alumni @jeffdonahue @jiayq @OriolVinyalsML @judyfhoffman , Ning Zhang, Eric Tzeng, and @trevordarrell for receiving the test of time award for the DeCAF paper at #ICML2024 this week in Vienna. https://
icml.cc/virtual/2024/3
8004
… -
Building applications with Transformer agents and Llama 3.1
By
–
I made it using Transformers agents. You can find all the code under "files" in the Space, in file "
http://
app.py"
It uses llama3.1-70B, which would be too heavy to run on a consumer machine, but you can use our API as I did in the space! -
Interactive LLM Decoding Strategies and Sampling Effects
By
–
Interactive LLM decoding strategies LLM sampling is a cool app that shows the effect of temperature, top K sampling, top P sampling, etc. on the probability distribution outputted by an LLM. LLM sampling: https://
artefact2.github.io/llm-sampling/i
ndex.xhtml
… -
Anticipation for Geometric Deep Learning Update at GRaM Workshop ICML
By
–
Looking forward to getting a full update on geometric DL at the GRaM workshop at ICML.
-
Advanced Retrieval-Augmented Generation Techniques Explained
By
–
Zain Hasan – Advanced Retrieval-Augmented Generation Techniques https://
bit.ly/3XSK0o8
#AI #MachineLearning #DeepLearning #LLMs #DataScience -
Project knowledge base limits constrained by LLM context windows
By
–
Currently the size of the project knowledge based is capped by the max context window Hopefully we will get a bigger one soon