Update prompts/README.md
Browse files- prompts/README.md +2 -0
prompts/README.md
CHANGED
|
@@ -1,6 +1,8 @@
|
|
| 1 |
EVERYTHING is synthetic - generated from danbooru, gelbooru, rule34, and e621 tag lists based on frequency and common use utilization meant to teach behavioral trajectory learning between T5_XL and CLIP_L + CLIP_G from illustrious.
|
| 2 |
|
| 3 |
# as of step 9000
|
|
|
|
|
|
|
| 4 |
The current state IS USABLE but it's not reliable yet with the T5. It's essentially formatted for the 100k and it's being fed that now at a smaller batch size and a new optimizer.
|
| 5 |
|
| 6 |
I expect the valuations to teach very useful information about counting people, and then the complexity for interactions can be solidified with smaller batch sizes for finetuning.
|
|
|
|
| 1 |
EVERYTHING is synthetic - generated from danbooru, gelbooru, rule34, and e621 tag lists based on frequency and common use utilization meant to teach behavioral trajectory learning between T5_XL and CLIP_L + CLIP_G from illustrious.
|
| 2 |
|
| 3 |
# as of step 9000
|
| 4 |
+
128 batch size so roughly 1,152,000 samples so far. Still have a long way to go.
|
| 5 |
+
|
| 6 |
The current state IS USABLE but it's not reliable yet with the T5. It's essentially formatted for the 100k and it's being fed that now at a smaller batch size and a new optimizer.
|
| 7 |
|
| 8 |
I expect the valuations to teach very useful information about counting people, and then the complexity for interactions can be solidified with smaller batch sizes for finetuning.
|