27 lines
424 B
Markdown
27 lines
424 B
Markdown
|
|
## Reproduce the paper results
|
|
|
|
### Download the dataset
|
|
|
|
```bash
|
|
huggingface-cli download --repo-type dataset PeterJinGo/nq_hotpotqa_train --local-dir $WORK_DIR/data/hotpot_qa
|
|
```
|
|
|
|
### Run PPO training
|
|
```bash
|
|
bash train_ppo.sh
|
|
```
|
|
|
|
|
|
### Run GRPO training
|
|
```bash
|
|
bash train_ppo.sh
|
|
```
|
|
|
|
### Run evaluation
|
|
```bash
|
|
bash evaluate.sh
|
|
```
|
|
|
|
You can change ```$BASE_MODEL``` to the path of the model you would like to evaluate.
|