Update README.md
Browse files
README.md
CHANGED
|
@@ -1,41 +1,46 @@
|
|
| 1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
|
| 3 |
## Training and evaluation data
|
| 4 |
### Training data
|
| 5 |
-
NaturalQuestions
|
|
|
|
|
|
|
| 6 |
|
| 7 |
### Evaluation data
|
| 8 |
-
NaturalQuestions
|
|
|
|
|
|
|
|
|
|
|
|
|
| 9 |
|
| 10 |
|
| 11 |
## Training hyperparameters
|
| 12 |
The following hyperparameters were used during training:
|
| 13 |
### SFT
|
| 14 |
-
learning_rate: 1e-6
|
| 15 |
-
train_batch_size: 256
|
| 16 |
-
train_micro_batch_size: 16
|
| 17 |
-
optimizer: Use adamw_torch with betas=(0.9,0.95)
|
| 18 |
-
lr_scheduler: cosine
|
| 19 |
-
num_epochs: 8.0
|
| 20 |
|
| 21 |
## RL
|
| 22 |
-
algorithm: dapo
|
| 23 |
-
learning_rate: 1e-7
|
| 24 |
-
lr_warmup_steps: 10
|
| 25 |
-
num_epochs: 1.0
|
| 26 |
|
| 27 |
## Framework versions
|
| 28 |
-
Transformers 4.57.6
|
| 29 |
-
Pytorch 2.4.0
|
| 30 |
-
Vllm 0.6.3
|
| 31 |
-
flash-attn 2.7.4.post1
|
| 32 |
|
| 33 |
-
---
|
| 34 |
-
license: apache-2.0
|
| 35 |
-
metrics:
|
| 36 |
-
- exact_match
|
| 37 |
-
- f1
|
| 38 |
-
- rouge
|
| 39 |
-
tags:
|
| 40 |
-
- GRIP
|
| 41 |
-
---
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
metrics:
|
| 4 |
+
- exact_match
|
| 5 |
+
- f1
|
| 6 |
+
- rouge
|
| 7 |
+
tags:
|
| 8 |
+
- GRIP
|
| 9 |
+
---
|
| 10 |
|
| 11 |
## Training and evaluation data
|
| 12 |
### Training data
|
| 13 |
+
- NaturalQuestions
|
| 14 |
+
- TriviaQA
|
| 15 |
+
- WebQuestions
|
| 16 |
|
| 17 |
### Evaluation data
|
| 18 |
+
- NaturalQuestions
|
| 19 |
+
- TriviaQA
|
| 20 |
+
- WebQuestions
|
| 21 |
+
- HotpotQA
|
| 22 |
+
- PopQA
|
| 23 |
|
| 24 |
|
| 25 |
## Training hyperparameters
|
| 26 |
The following hyperparameters were used during training:
|
| 27 |
### SFT
|
| 28 |
+
- learning_rate: 1e-6
|
| 29 |
+
- train_batch_size: 256
|
| 30 |
+
- train_micro_batch_size: 16
|
| 31 |
+
- optimizer: Use adamw_torch with betas=(0.9,0.95)
|
| 32 |
+
- lr_scheduler: cosine
|
| 33 |
+
- num_epochs: 8.0
|
| 34 |
|
| 35 |
## RL
|
| 36 |
+
- algorithm: dapo
|
| 37 |
+
- learning_rate: 1e-7
|
| 38 |
+
- lr_warmup_steps: 10
|
| 39 |
+
- num_epochs: 1.0
|
| 40 |
|
| 41 |
## Framework versions
|
| 42 |
+
- Transformers 4.57.6
|
| 43 |
+
- Pytorch 2.4.0
|
| 44 |
+
- Vllm 0.6.3
|
| 45 |
+
- flash-attn 2.7.4.post1
|
| 46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|