tangger/Search-R1
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

response length include retrieval info

Browse Source
This commit is contained in:
PeterGriffinJin
2025-03-19 00:36:21 +00:00
parent 50cedb2c00
commit 8c7f04ca45
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
verl/trainer/ppo/ray_trainer.py
View File

@@ -163,8 +163,8 @@ def reduce_metrics(metrics: dict):
def _compute_response_info(batch): def _compute_response_info(batch):
response_length = batch.batch['responses'].shape[-1] response_length = batch.batch['responses'].shape[-1]
prompt_mask = batch.batch['info_mask'][:, :-response_length] prompt_mask = batch.batch['attention_mask'][:, :-response_length]
response_mask = batch.batch['info_mask'][:, -response_length:] response_mask = batch.batch['attention_mask'][:, -response_length:]
prompt_length = prompt_mask.sum(-1).float() prompt_length = prompt_mask.sum(-1).float()
response_length = response_mask.sum(-1).float() # (batch_size,) response_length = response_mask.sum(-1).float() # (batch_size,)