Commit
•
06607c1
1
Parent(s):
ab54f6e
Update README.md
Browse files
README.md
CHANGED
|
@@ -12,7 +12,22 @@ license: llama3
|
|
| 12 |
|
| 13 |
## Training Details
|
| 14 |
|
| 15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
## Limitations
|
| 18 |
|
|
|
|
| 12 |
|
| 13 |
## Training Details
|
| 14 |
|
| 15 |
+
Hardware: 1x RTX 4090
|
| 16 |
+
Duration: ~30 hours in total (~2 hours for first phase and ~28 hours for second phase)
|
| 17 |
+
|
| 18 |
+
### Hyperparameters
|
| 19 |
+
|
| 20 |
+
Adapter: QLoRA
|
| 21 |
+
Precision: 4 bit
|
| 22 |
+
Optimizer: adamw_bnb_8bit
|
| 23 |
+
LoRA Rank: 256
|
| 24 |
+
LoRA Alpha: 256
|
| 25 |
+
Learning Rate: 1e-5
|
| 26 |
+
Context Length: 4096 tokens
|
| 27 |
+
Batch Size: 1
|
| 28 |
+
Gradient Accumulation Steps: 1
|
| 29 |
+
Sample Packing: Off for first phase, on for second phase
|
| 30 |
+
Epochs: 2
|
| 31 |
|
| 32 |
## Limitations
|
| 33 |
|