Note: For fine-tuning the Pile model, change 1e-15 to 1e-9 in https://github.com/BlinkDL/RWKV-LM/blob/main/RWKV-v2-RNN/src/model.py and https://github.com/BlinkDL/RWKV-LM/blob/main/RWKV-v2-RNN/src/model_run.py and probably you need other changes as well. You can compare the output with the latest code ( https://github.com/BlinkDL/RWKV-v2-RNN-Pile ) to verify it.
I usually fine-tune with 4e-5 lr, and decay to 1e-5 when it plateaus.
## How it works
RWKV is inspired by Apple's AFT (https://arxiv.org/abs/2105.14103).