Update README.md
Browse files
README.md
CHANGED
|
@@ -3,13 +3,15 @@ Training Details:
|
|
| 3 |
Trained at 8K Context -> Expanded to 32K Context due to context extension with PoSE training.
|
| 4 |
|
| 5 |
Dataset Modifications:
|
| 6 |
-
- Further Cleaned up Roleplaying Samples
|
| 7 |
- Removed Low Quality Samples from Manual Check
|
| 8 |
-
- More Creative Writing Samples
|
| 9 |
- Remade and Refined Detailed Instruct Data
|
| 10 |
|
| 11 |
Needle in a Haystack Results:
|
|
|
|
| 12 |
|
|
|
|
| 13 |
|
| 14 |
```
|
| 15 |
sequence_len: 8192
|
|
|
|
| 3 |
Trained at 8K Context -> Expanded to 32K Context due to context extension with PoSE training.
|
| 4 |
|
| 5 |
Dataset Modifications:
|
| 6 |
+
- Further Cleaned up Roleplaying Samples -> Quality Check
|
| 7 |
- Removed Low Quality Samples from Manual Check
|
| 8 |
+
- More Creative Writing Samples -> 2x
|
| 9 |
- Remade and Refined Detailed Instruct Data
|
| 10 |
|
| 11 |
Needle in a Haystack Results:
|
| 12 |
+
|
| 13 |
|
| 14 |
+
Coherent at 32K Context. Not as good as a natively trained 32K model, but much better than regular rope scaling.
|
| 15 |
|
| 16 |
```
|
| 17 |
sequence_len: 8192
|