Update README.md
Browse files
README.md
CHANGED
|
@@ -4,13 +4,20 @@ language:
|
|
| 4 |
- en
|
| 5 |
---
|
| 6 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
Fimbulvetr-v2 but extended to 16K with PoSE. A sane context value would be ~12K before it degrades.
|
| 8 |
-
|
| 9 |
-
|
|
|
|
| 10 |
|
| 11 |
Notes:
|
| 12 |
<br> \- I noticed peoplle having bad issues with quants. Be it GGUF or others, at 8 bit or less. Kind of a weird issue? I had little to no issues during testing at the full precision
|
| 13 |
<br> \- Slightly different results from base Fimbulvetr-v2, but during my tests they are similar enough. The vibes are still there.
|
| 14 |
<br> \- Formatting issues happen rarely. Sometimes. A reroll / regenerate fixes it from tests.
|
|
|
|
|
|
|
| 15 |
|
| 16 |
|
|
|
|
| 4 |
- en
|
| 5 |
---
|
| 6 |
|
| 7 |
+
Trained with compute from Backyard.ai | Thanks to them and @dynafire for helping me out.
|
| 8 |
+
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
Fimbulvetr-v2 but extended to 16K with PoSE. A sane context value would be ~12K before it degrades.
|
| 12 |
+
|
| 13 |
+
Note:
|
| 14 |
+
<br> \- I left Rope Theta at 10K for this train, instead of expanding it like with Stheno 3.3. Solar did not play will with extended theta, grad norm / loss values went parabolic or plunged from 10000+ down. Unreliable pretty much, unlike Stheno 3.3's training run.
|
| 15 |
|
| 16 |
Notes:
|
| 17 |
<br> \- I noticed peoplle having bad issues with quants. Be it GGUF or others, at 8 bit or less. Kind of a weird issue? I had little to no issues during testing at the full precision
|
| 18 |
<br> \- Slightly different results from base Fimbulvetr-v2, but during my tests they are similar enough. The vibes are still there.
|
| 19 |
<br> \- Formatting issues happen rarely. Sometimes. A reroll / regenerate fixes it from tests.
|
| 20 |
+
<br> \- I get consistent and reliable answers at ~11K context fine.
|
| 21 |
+
<br> \- Still coherent at up to 16K though! Just works not that well.
|
| 22 |
|
| 23 |
|