One of the important things to be considered in few-shot classification with discrete text prompting is, which shots should be selected to yield a good performance. There are some studies to figure out which (e.g. random selection, closest selection by top-k similarities), but these have some problems to be resolved. Motivated by RLPrompt that attaches MLP module to the gradient-fixed language model and trains it with reinforcement learning, we propose a way to select the shots by a policy constructed by RL.
Just execute train.py with some arguments. You will find these definitions in ArgumentParser at the bottom of the codes. Here is an example.
python train.py --dataset super_glue,cb --shot_num 2 --batch_size 4 --lr 1e-3 --weight_decay 0.0 --topk 5 --tv_split_ratio 0.0
If you want to test it, execute test.py. For instance,
python test.py --dataset super_glue,cb --batch_size 4
The test set accuracy of random and ours is measured with the setting which showed the best performance on the validation set.
| cb | random | closest | ours |
|---|---|---|---|
| 2 shots | 32.1 | 46.4 | 64.3 |
| 3 shots | 35.7 | 53.6 | 75 |
| 4 shots | 53.6 | 57.1 | 46.4 |
| boolq | random | closest | ours |
|---|---|---|---|
| 2 shots | 48.3 | 53.3 | 55 |
| 3 shots | 51.7 | 53.8 | 60.4 |
| 4 shots | 57.1 | 54.3 | 60.2 |
| sst2 | random | closest | ours |
|---|---|---|---|
| 2 shots | 69.7 | 77.3 | 82.8 |
| 3 shots | 73.4 | 75.9 | 82.6 |
| 4 shots | 76.4 | 73.2 | 81.2 |