open-thought/reasoning-gym
1
0
Fork 0
mirror of https://github.com/open-thought/reasoning-gym.git synced 2026-04-30 17:40:45 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

updated correctness score func

Browse source
This commit is contained in:
joesharratt1229 2025-03-26 19:54:37 +00:00
parent f50c7221ac
commit cc0bacd8e1
1 changed files with 5 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

9
training/rewards/reward.py
View file

@ -33,9 +33,9 @@ reward_registry = RewardRegistry()
@reward_registry.register("cosine")
def cosine_scaled_reward(solution_str, scaling_factor, **kwargs):
"""Reward function that scales based on completion length using a cosine schedule."""
min_value_wrong = -1.0
max_value_wrong = -0.5
min_value_correct = 0.5
min_value_wrong = 0
max_value_wrong = 0.7
min_value_correct = 0.95
max_value_correct = 1.0
max_len = 1000
@ -77,8 +77,9 @@ def compute_format_reward(solution_str: str, scaling_factor: float = 0.2, **kwar
@reward_registry.register("length")
def length_reward(solution_str, correctness_score, scaling_factor, **kwargs):
def length_reward(solution_str, scaling_factor, **kwargs):
"""Reward length appropriately based on correctness."""
correctness_score = kwargs.get("correctness_score", 0.0)
epsilon = 1e-6
max_score = kwargs.get("max_score", 1.0)
max_output_length = kwargs.get("max_output_length", 1024)