End of training
Browse files- README.md +12 -20
- adapter_model.bin +1 -1
README.md
CHANGED
@@ -67,7 +67,7 @@ wandb_name: test
|
|
67 |
gradient_accumulation_steps: 4
|
68 |
micro_batch_size: 2 # was 16
|
69 |
eval_batch_size: 2 # was 16
|
70 |
-
num_epochs:
|
71 |
optimizer: paged_adamw_32bit
|
72 |
lr_scheduler: cosine
|
73 |
learning_rate: 0.0002
|
@@ -104,12 +104,12 @@ special_tokens:
|
|
104 |
|
105 |
</details><br>
|
106 |
|
107 |
-
[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="200" height="32"/>](https://wandb.ai/sncds/ocr-ft/runs/
|
108 |
# llama-3-8b-ocr-correction
|
109 |
|
110 |
This model is a fine-tuned version of [meta-llama/Meta-Llama-3-8B](https://huggingface.co/meta-llama/Meta-Llama-3-8B) on the None dataset.
|
111 |
It achieves the following results on the evaluation set:
|
112 |
-
- Loss: 0.
|
113 |
|
114 |
## Model description
|
115 |
|
@@ -137,29 +137,21 @@ The following hyperparameters were used during training:
|
|
137 |
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
138 |
- lr_scheduler_type: cosine
|
139 |
- lr_scheduler_warmup_steps: 10
|
140 |
-
- num_epochs:
|
141 |
|
142 |
### Training results
|
143 |
|
144 |
| Training Loss | Epoch | Step | Validation Loss |
|
145 |
|:-------------:|:------:|:----:|:---------------:|
|
146 |
| 0.6611 | 0.0165 | 1 | 0.6229 |
|
147 |
-
| 0.
|
148 |
-
| 0.
|
149 |
-
| 0.
|
150 |
-
| 0.
|
151 |
-
| 0.
|
152 |
-
| 0.
|
153 |
-
| 0.
|
154 |
-
| 0.
|
155 |
-
| 0.0622 | 2.1811 | 135 | 0.1879 |
|
156 |
-
| 0.0674 | 2.4280 | 150 | 0.1868 |
|
157 |
-
| 0.0715 | 2.6749 | 165 | 0.1876 |
|
158 |
-
| 0.0567 | 2.9218 | 180 | 0.1851 |
|
159 |
-
| 0.0445 | 3.1481 | 195 | 0.1928 |
|
160 |
-
| 0.0419 | 3.3951 | 210 | 0.2017 |
|
161 |
-
| 0.0371 | 3.6420 | 225 | 0.2021 |
|
162 |
-
| 0.0382 | 3.8889 | 240 | 0.2022 |
|
163 |
|
164 |
|
165 |
### Framework versions
|
|
|
67 |
gradient_accumulation_steps: 4
|
68 |
micro_batch_size: 2 # was 16
|
69 |
eval_batch_size: 2 # was 16
|
70 |
+
num_epochs: 2
|
71 |
optimizer: paged_adamw_32bit
|
72 |
lr_scheduler: cosine
|
73 |
learning_rate: 0.0002
|
|
|
104 |
|
105 |
</details><br>
|
106 |
|
107 |
+
[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="200" height="32"/>](https://wandb.ai/sncds/ocr-ft/runs/m4qbupk5)
|
108 |
# llama-3-8b-ocr-correction
|
109 |
|
110 |
This model is a fine-tuned version of [meta-llama/Meta-Llama-3-8B](https://huggingface.co/meta-llama/Meta-Llama-3-8B) on the None dataset.
|
111 |
It achieves the following results on the evaluation set:
|
112 |
+
- Loss: 0.1742
|
113 |
|
114 |
## Model description
|
115 |
|
|
|
137 |
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
138 |
- lr_scheduler_type: cosine
|
139 |
- lr_scheduler_warmup_steps: 10
|
140 |
+
- num_epochs: 2
|
141 |
|
142 |
### Training results
|
143 |
|
144 |
| Training Loss | Epoch | Step | Validation Loss |
|
145 |
|:-------------:|:------:|:----:|:---------------:|
|
146 |
| 0.6611 | 0.0165 | 1 | 0.6229 |
|
147 |
+
| 0.3149 | 0.2469 | 15 | 0.2870 |
|
148 |
+
| 0.2074 | 0.4938 | 30 | 0.2166 |
|
149 |
+
| 0.2211 | 0.7407 | 45 | 0.1937 |
|
150 |
+
| 0.195 | 0.9877 | 60 | 0.1825 |
|
151 |
+
| 0.1411 | 1.2140 | 75 | 0.1787 |
|
152 |
+
| 0.1348 | 1.4609 | 90 | 0.1760 |
|
153 |
+
| 0.1479 | 1.7078 | 105 | 0.1743 |
|
154 |
+
| 0.1413 | 1.9547 | 120 | 0.1742 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
|
156 |
|
157 |
### Framework versions
|
adapter_model.bin
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 167934026
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c64465bb2211b47808dc809512a591f6ada32a06c95e2e5ae6b3bef6b9622301
|
3 |
size 167934026
|