Awesome! You mentioned LLaMA 2 versions are coming soon, what’s the base model here?
@rasbt
-
Less Data Better Results Finetuning LLMs with Trimmed Datasets
By
–
Next to LIMA, this is another interesting paper highlighting that more data is not always better when finetuning LLMs: https://
arxiv.org/abs/2307.08701 Trimming the orig 52k Alpaca dataset to 9k can improve the performance when finetuning 7B and 13B parameter LLMs. -
Brief CNN testing experience, focus shifted to LLM work
By
–
Phew, it's been a while. I think I only briefly tried it on some basic CNNs, and it worked fine out of the box. But afaik it's mainly for computer vision applications (have been super focused on LLM work recently)
-
Figure 1 Users: Stop Requesting Permission for Each Use
By
–
"Hey, you >50k people using Figure 1 in your blog articles and lecture slides, pls don't email us for permission each time."
-
Multiprocessing faster than tested alternative, future no-GIL Python promising
By
–
Ha, tried that, but it was slower than multiprocessing. But yeah, with No Gil future Python it might be nice!
-
NN+Gzip optimization: caching and parallel processing improvements
By
–
Update on the NN+Gzip front. Added
1) a caching mechanism, which speeds it up 25%.
2) Parallel processing to speed it up 400% (on a MB Air). Love this method. It's a nice test case for teaching multiprocessing! Code: https://
github.com/rasbt/nn_plus_
gzip/blob/main/1_2_caching-multiprocessing.py
… -
Suggestions for More Interesting Model Training Progression Sequences
By
–
Ok fair, it would have been of course more interesting if they did SFT v1 -> SFT v2 -> SFT v3 -> … SFT v1 -> SFT v2 -> RLHF v1 -> … RLHF v1 -> RLHF v2 -> RLHF v2 -> …
-
DPO vs RLHF: Genuine Competition or Surface-Level Improvement?
By
–
Yeah exactly. I wonder if DPO genuinely competes with RLHF or if the models only looks good on the surface but are worse under closer inspection (like with imitation models)
-
Human Evaluations Included in Upper Right Instead of GPT-4
By
–
I included human evaluations in the upper right as opposed to GPT-4 evaluations