Re:Pile-T5
by gwern on Monday April 15, 2024 @12:32PM
(#40042729)
Now that's a blast from the past - T5! I always thought it was underused, but it's also from so long ago now, and even Google has moved past it to UL2 etc AFAIK. What's the use-case here for reproducing it? Aren't there already many good code models?
[ Parent
| on HN ]
my group is currently working on a T5 model (and tokenizer) for html, as there are very few (if any) tokenizers that work well with HTML!
You can try using GPT4's tokenizer on your own HTML inputs below [1] ... there's definitely room for improvement!
Blast from the past? Like 2 years ago?
T5 was released in 2019.
Isn't that two years ago?
Re:Pile-T5
by littlestymaar on Monday April 15, 2024 @01:50PM
(#40043585)
Two years already? This locked-down year of 2020 really messes up with me perception of time.
Gather round, children, and let Grandpa tell you a tale from way back in the day, when we thought 24GB was a lot of VRAM...
Re:Pile-T5
by bevekspldnw on Monday April 15, 2024 @01:29PM
(#40043350)
My nearly new 48GB A6000 is basically an ancient artifact relative to the exploding size of MoE requirements.
I figured it would be interesting to see scale and for the image models
A lot of embedding models are built on top of T5's encoder, this offers a new option
The modularity of the enc-dec approach is useful - you can insert additional models in between (e.g. A diffusion model), you can use different encoders for different modalities, etc
Re:Pile-T5
by jakderrida on Wednesday April 17, 2024 @06:31PM
(#40070766)
I've downloaded the data for every model on Huggingface and, if nothing else, flan-t5-xxl is the largest encoder (discriminant) model in terms of total file sizes. To be clear T5 is an encoder-decoder model. But the shift in focus to decoder-only has made it so encoder and encoder-decoder models have been, for lack of a better word, neglected, in terms of training focus.