Reward Functio Choosing Error

[forthelostthings]

I studied that in the stage 2 RL, the data source name is codeio-forward-incomplete-depth2. But for your default setting the mp reward class would raise not impemented error for this data source name. Is there anything that I should do before running stage 2 RL to avoid this happen?

[chenweize1998]

Hi! Sorry for that. That was actually a bug we introduced during code cleanup. We’ve fixed the conditional branch in the reward function selection. Please give it another try.

[forthelostthings]

Hi! Sorry for that. That was actually a bug we introduced during code cleanup. We’ve fixed the conditional branch in the reward function selection. Please give it another try.

Thanks for your response! Now I am doing the stage 2 level 2 rft baseline, and there is a "FileNotFoundError: Unable to find '/username/RL-Compositionality/data/string_task/stage2_leve2/train.parquet'", may I ask whether it is stage_level2 or stage2_level2_rft?

Apart from that, I found that save_path "SAVE_PATH=data/string_task/stage2_rft_level2/rollout.parquet" is for the gen_path of string_manipulation_sft.py, and dataset = load_dataset("parquet", data_files=gen_path)['train'](strxxx.py line59) would result in another file not found. May I ask what should be the dataset?