modified by Mr_KrzYch00 for an innovative website front-end support.
The AI tries to predict the continuation of the text that has been typed in the INPUT box
Generated samples are browser session locked and visible only for that session
1558M model generates 1 sample at a time with 10-token self-input-accumulating micro-samples for a "maximum 1024 tokens" sliding-window-like effect
Micro-batches undergo post-processed, allowing more control over the re-feeding micro-batch to input mechanism for outputs longer than 20 tokens
All available AI parameters are adjustable: Temperature, TOP_K and TOP_P, with TOP_P having an additional multi-range support
The <|endoftext|> token is
supported in the INPUT box and
correctly passed to and understood by the AI
The amount of tokens detected in the INPUT box is displayed
as the AI sees them (functionality integrated within the front-end)
Simple tooltip functionality is provided for clarification about each of the functions this website provides
Virtual windows on this site can be repositioned and resized freely within the browser window
Powered by Nvidia GTX 3060 12GB (barely fits; sampling ~10T/4.5s due to inability to utilize all CUDA cores [auto-regressive decoding])
Old projects and stuff here
Please wait . . .
▲ Advanced Parameters ▲
TOP_P List (0 ~ 1) |
Remove repeated text
▲ Input TEXT ▲
Send to GPT-2 1558M
▲ Output SAMPLES ▲
Generated samples list (click buttons below to open AI reply)
This is a research site, do not abuse!
The administrator is not responsible for any input user writes as well as the output the algorithm produces.