github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-06-20 12:43:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix reward score

Browse Source
This commit is contained in:
Tong Li 2025-03-11 10:17:32 +08:00
parent 71a0181fce
commit abca66e69f
1 changed files with 6 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
applications/ColossalChat/coati/distributed/reward/reward_fn.py
View File

@ -4,6 +4,8 @@ from .reward_utils import extract_solution, validate_response_structure
def math_reward_fn(input_ids, gt_answer, response_idx, **kwargs):
format_score = 1.0
acc_score = 9.0
tokenizer = kwargs["tokenizer"]
reward = torch.tensor(0.0)
format_reward = torch.tensor(0.0)
@ -20,16 +22,16 @@ def math_reward_fn(input_ids, gt_answer, response_idx, **kwargs):
# Check format accuracy
if format_valid:
format_reward += 1.0
reward += 1.0
format_reward += format_score
reward += format_score
# Check answer accuracy
if (
final_answer is not None
and gt_answer.strip().replace(" ", "").lower() == final_answer.strip().replace(" ", "").lower()
):
acc_reward += 5.0
reward += 5.0
acc_reward += acc_score
reward += acc_score
return torch.tensor([reward, format_reward, acc_reward]).to(input_ids.device)