Update README.md
Browse files
README.md
CHANGED
|
@@ -12,37 +12,12 @@ should probably proofread and complete it, then remove this comment. -->
|
|
| 12 |
|
| 13 |
# ProGemma2
|
| 14 |
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
## Model description
|
| 18 |
-
|
| 19 |
-
More information needed
|
| 20 |
-
|
| 21 |
-
## Intended uses & limitations
|
| 22 |
-
|
| 23 |
-
More information needed
|
| 24 |
-
|
| 25 |
-
## Training and evaluation data
|
| 26 |
-
|
| 27 |
-
More information needed
|
| 28 |
-
|
| 29 |
-
## Training procedure
|
| 30 |
-
|
| 31 |
-
### Training hyperparameters
|
| 32 |
-
|
| 33 |
-
The following hyperparameters were used during training:
|
| 34 |
-
- learning_rate: 0.001
|
| 35 |
-
- train_batch_size: 2
|
| 36 |
-
- eval_batch_size: 8
|
| 37 |
-
- seed: 42
|
| 38 |
-
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
| 39 |
-
- lr_scheduler_type: linear
|
| 40 |
-
- lr_scheduler_warmup_ratio: 0.4
|
| 41 |
-
- training_steps: 3500
|
| 42 |
-
|
| 43 |
-
### Training results
|
| 44 |
|
|
|
|
|
|
|
| 45 |
|
|
|
|
| 46 |
|
| 47 |
### Framework versions
|
| 48 |
|
|
|
|
| 12 |
|
| 13 |
# ProGemma2
|
| 14 |
|
| 15 |
+
Identical to JuIm/ProGemma, save for 2 details:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
+
1. The model is slightly larger at 336M parameters vs 225M
|
| 18 |
+
2. The training rate is 1e-3
|
| 19 |
|
| 20 |
+
Current training loss is 2.04, with only 20% of the training data being used thus far (1st epoch), which is a marked improvement versus the original ProGemma's training loss at this point in the dataset.
|
| 21 |
|
| 22 |
### Framework versions
|
| 23 |
|