Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Jam-CGPT
|
| 2 |
+
|
| 3 |
+
Jam-CGPT is a GPT2-like model that follows [jam](https://huggingface.co/apcl/jam)'s pretraining procedure to pretrain models ranging from 38 million to 350 million parameters and finetuning with comments generated by GPT-3.5 and data size ranging from 170k to 2.15m.
|
| 4 |
+
|
| 5 |
+
## Jam-CGPT Training Details
|
| 6 |
+
- We follow [jam](https://huggingface.co/apcl/jam)'s pretraining procedure and use the same data to pretrain our 38m, 110m and 350m parameters models.
|
| 7 |
+
- We finetune our Jam-CGPT with the summaries generated by GPT-3.5 and 4 different dataset size [Jam-CGPT dataset](https://huggingface.co/datasets/apcl/Jam-CGPT).
|
| 8 |
+
- We finetune our models for 3 epochs.
|
| 9 |
+
- Our [GitHub repo](https://github.com/apcl-research/Jam-CGPT) contains the code for reproduction using the same [data](https://huggingface.co/datasets/apcl/Jam-CGPT).
|
| 10 |
+
|
| 11 |
+
## Jam-CGPT 38 million parameters model
|
| 12 |
+
| Hyperparameter | Description | Value |
|
| 13 |
+
| ----------- | ----------- |------------|
|
| 14 |
+
|e | embedding dimensions | 512 |
|
| 15 |
+
|L | number of layers | 4 |
|
| 16 |
+
|h | attention heads | 4 |
|
| 17 |
+
|c | block size / context length | 256 |
|
| 18 |
+
|b | batch size | 64 |
|
| 19 |
+
|a | accumulation steps | 2 |
|
| 20 |
+
|d | dropout | 0.20 |
|
| 21 |
+
|r | learning rate | 3e-5 |
|
| 22 |
+
|y | weight decay | 1e-5 |
|
| 23 |
+
|
| 24 |
+
## Jam-CGPT 110 million parameters model
|
| 25 |
+
| Hyperparameter | Description | Value |
|
| 26 |
+
| ----------- | ----------- |------------|
|
| 27 |
+
|e | embedding dimensions | 768 |
|
| 28 |
+
|L | number of layers | 10|
|
| 29 |
+
|h | attention heads | 8 |
|
| 30 |
+
|c | block size / context length | 256 |
|
| 31 |
+
|b | batch size | 32 |
|
| 32 |
+
|a | accumulation steps | 4 |
|
| 33 |
+
|d | dropout | 0.20 |
|
| 34 |
+
|r | learning rate | 3e-5 |
|
| 35 |
+
|y | weight decay | 1e-5 |
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
## Jam-CGPT 350 million parameters model
|
| 39 |
+
| Hyperparameter | Description | Value |
|
| 40 |
+
| ----------- | ----------- |------------|
|
| 41 |
+
|e | embedding dimensions | 1024 |
|
| 42 |
+
|L | number of layers | 24 |
|
| 43 |
+
|h | attention heads | 16 |
|
| 44 |
+
|c | block size / context length | 256 |
|
| 45 |
+
|b | batch size | 4 |
|
| 46 |
+
|a | accumulation steps | 32 |
|
| 47 |
+
|d | dropout | 0.20 |
|
| 48 |
+
|r | learning rate | 3e-5 |
|
| 49 |
+
|y | weight decay | 1e-5 |
|
| 50 |
+
|
| 51 |
+
- Note that you can adjust the batch size and accumulation steps based on your GPU memory. But, the batch size * accumulation steps should be 128.
|
| 52 |
+
- If you finetune your models with multiple GPUs, you can turn down accumulation steps. For example, if you finetune with 2 GPUs, you will need to half the accumulation steps.
|
| 53 |
+
|