Update README.md
Browse files
README.md
CHANGED
|
@@ -4,6 +4,8 @@ license: apache-2.0
|
|
| 4 |
# ArliAI-RPMax-12B-v1.2
|
| 5 |
=====================================
|
| 6 |
|
|
|
|
|
|
|
| 7 |
## RPMax Series Overview
|
| 8 |
|
| 9 |
| [2B](https://huggingface.co/ArliAI/Gemma-2-2B-ArliAI-RPMax-v1.1) |
|
|
@@ -38,7 +40,7 @@ v1.2 update completely removes non-creative/RP examples in the dataset and is al
|
|
| 38 |
* **Sequence Length**: 8192
|
| 39 |
* **Training Duration**: Approximately 2 days on 2x3090Ti
|
| 40 |
* **Epochs**: 1 epoch training for minimized repetition sickness
|
| 41 |
-
* **
|
| 42 |
* **Learning Rate**: 0.00001
|
| 43 |
* **Gradient accumulation**: Very low 32 for better learning.
|
| 44 |
|
|
|
|
| 4 |
# ArliAI-RPMax-12B-v1.2
|
| 5 |
=====================================
|
| 6 |
|
| 7 |
+
## UPDATE: Merged it wrongly to base after LORA training. Working to reupload. The 8B version meanwhile is working fine.
|
| 8 |
+
|
| 9 |
## RPMax Series Overview
|
| 10 |
|
| 11 |
| [2B](https://huggingface.co/ArliAI/Gemma-2-2B-ArliAI-RPMax-v1.1) |
|
|
|
|
| 40 |
* **Sequence Length**: 8192
|
| 41 |
* **Training Duration**: Approximately 2 days on 2x3090Ti
|
| 42 |
* **Epochs**: 1 epoch training for minimized repetition sickness
|
| 43 |
+
* **LORA**: 64-rank 128-alpha, resulting in ~2% trainable weights
|
| 44 |
* **Learning Rate**: 0.00001
|
| 45 |
* **Gradient accumulation**: Very low 32 for better learning.
|
| 46 |
|