Add reward functions and registry
Browse files- src/rewards.py +156 -0
src/rewards.py
ADDED
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
from typing import List, Dict, Optional, Callable
|
3 |
+
import numpy as np
|
4 |
+
|
5 |
+
def format_reward(completions, **kwargs):
|
6 |
+
"""Reward function that checks if the code is enclosed within <code> and </code> tags,
|
7 |
+
and the final answer is enclosed within <answer> and </answer> tags."""
|
8 |
+
pattern = r".*?<code>.*?</code>.*?<answer>.*?</answer>.*?"
|
9 |
+
completion_contents = [completion[0]["content"] for completion in completions]
|
10 |
+
matches = [re.search(pattern, content, re.DOTALL | re.MULTILINE) is not None for content in completion_contents]
|
11 |
+
return [1.0 if match else 0.0 for match in matches]
|
12 |
+
|
13 |
+
def accuracy_reward(completions: list[list[dict[str, str]]], correct_answers: list[str], **kwargs) -> list[Optional[float]]:
|
14 |
+
"""Reward function that checks if the completion's answer matches the ground truth."""
|
15 |
+
contents = [completion[0]["content"] for completion in completions]
|
16 |
+
rewards = []
|
17 |
+
|
18 |
+
for content, correct_answer in zip(contents, correct_answers):
|
19 |
+
# Extract answer from the completion using regex
|
20 |
+
answer_match = re.search(r'<answer>\\boxed{(.*?)}</answer>', content, re.DOTALL)
|
21 |
+
if not answer_match:
|
22 |
+
rewards.append(0.0)
|
23 |
+
continue
|
24 |
+
|
25 |
+
extracted_answer = answer_match.group(1).strip()
|
26 |
+
|
27 |
+
# Check if the extracted answer matches the correct answer
|
28 |
+
# You might need a more sophisticated comparison for mathematical expressions
|
29 |
+
if extracted_answer == correct_answer:
|
30 |
+
rewards.append(1.0)
|
31 |
+
else:
|
32 |
+
rewards.append(0.0)
|
33 |
+
|
34 |
+
return rewards
|
35 |
+
|
36 |
+
def code_execution_reward(completions, **kwargs):
|
37 |
+
"""Reward function that checks if the code execution was successful."""
|
38 |
+
completion_contents = [completion[0]["content"] for completion in completions]
|
39 |
+
# Check for error patterns in interpreter output
|
40 |
+
error_patterns = [
|
41 |
+
r'<interpreter>.*?Error.*?</interpreter>',
|
42 |
+
r'<interpreter>.*?Exception.*?</interpreter>',
|
43 |
+
r'<interpreter>.*?Traceback.*?</interpreter>'
|
44 |
+
]
|
45 |
+
|
46 |
+
rewards = []
|
47 |
+
for content in completion_contents:
|
48 |
+
# Find all code-interpreter pairs
|
49 |
+
code_blocks = re.findall(r'<code>.*?</code>\s*<interpreter>(.*?)</interpreter>', content, re.DOTALL)
|
50 |
+
if not code_blocks:
|
51 |
+
rewards.append(0.0)
|
52 |
+
continue
|
53 |
+
|
54 |
+
# Check each interpreter output for errors
|
55 |
+
error_count = 0
|
56 |
+
for interpreter_output in code_blocks:
|
57 |
+
has_error = any(re.search(pattern, interpreter_output, re.DOTALL) for pattern in error_patterns)
|
58 |
+
if has_error:
|
59 |
+
error_count += 1
|
60 |
+
|
61 |
+
# Calculate success rate
|
62 |
+
if len(code_blocks) == 0:
|
63 |
+
rewards.append(0.0)
|
64 |
+
else:
|
65 |
+
success_rate = 1.0 - (error_count / len(code_blocks))
|
66 |
+
rewards.append(success_rate)
|
67 |
+
|
68 |
+
return rewards
|
69 |
+
|
70 |
+
def len_reward(completions, **kwargs):
|
71 |
+
"""Reward shorter completions to encourage efficiency."""
|
72 |
+
completion_contents = [completion[0]["content"] for completion in completions]
|
73 |
+
lengths = [len(content) for content in completion_contents]
|
74 |
+
|
75 |
+
# If all completions have the same length, return neutral rewards
|
76 |
+
if min(lengths) == max(lengths):
|
77 |
+
return [0.0] * len(completions)
|
78 |
+
|
79 |
+
# Normalize lengths to [0, 1] range and invert (shorter = higher reward)
|
80 |
+
normalized_lengths = [(length - min(lengths)) / (max(lengths) - min(lengths)) for length in lengths]
|
81 |
+
rewards = [1.0 - norm_length for norm_length in normalized_lengths]
|
82 |
+
|
83 |
+
# Scale to a smaller range to make this a secondary consideration
|
84 |
+
scaled_rewards = [0.2 * reward for reward in rewards]
|
85 |
+
|
86 |
+
return scaled_rewards
|
87 |
+
|
88 |
+
def code_ratio_reward(completions, **kwargs):
|
89 |
+
"""Reward appropriate code-to-text ratio."""
|
90 |
+
completion_contents = [completion[0]["content"] for completion in completions]
|
91 |
+
rewards = []
|
92 |
+
|
93 |
+
for content in completion_contents:
|
94 |
+
# Extract all code blocks
|
95 |
+
code_blocks = re.findall(r'<code>(.*?)</code>', content, re.DOTALL)
|
96 |
+
total_code_length = sum(len(code) for code in code_blocks)
|
97 |
+
total_length = len(content)
|
98 |
+
|
99 |
+
if total_length == 0:
|
100 |
+
rewards.append(0.0)
|
101 |
+
continue
|
102 |
+
|
103 |
+
code_ratio = total_code_length / total_length
|
104 |
+
|
105 |
+
# Reward an optimal ratio range (e.g., 0.2 to 0.4)
|
106 |
+
if 0.2 <= code_ratio <= 0.4:
|
107 |
+
rewards.append(0.3) # Full reward
|
108 |
+
elif 0.1 <= code_ratio < 0.2 or 0.4 < code_ratio <= 0.5:
|
109 |
+
rewards.append(0.2) # Partial reward
|
110 |
+
elif 0.05 <= code_ratio < 0.1 or 0.5 < code_ratio <= 0.6:
|
111 |
+
rewards.append(0.1) # Minimal reward
|
112 |
+
else:
|
113 |
+
rewards.append(0.0) # No reward
|
114 |
+
|
115 |
+
return rewards
|
116 |
+
|
117 |
+
def code_timing_reward(completions, **kwargs):
|
118 |
+
"""Reward for invoking code at appropriate points in the reasoning process."""
|
119 |
+
completion_contents = [completion[0]["content"] for completion in completions]
|
120 |
+
rewards = []
|
121 |
+
|
122 |
+
for content in completion_contents:
|
123 |
+
# Calculate relative position of first code block
|
124 |
+
first_code_pos = content.find('<code>')
|
125 |
+
if first_code_pos == -1:
|
126 |
+
rewards.append(0.0)
|
127 |
+
continue
|
128 |
+
|
129 |
+
relative_pos = first_code_pos / len(content)
|
130 |
+
|
131 |
+
# Reward early-to-middle code invocation (between 10% and 40% of the way through)
|
132 |
+
if 0.1 <= relative_pos <= 0.4:
|
133 |
+
rewards.append(0.3)
|
134 |
+
elif 0.05 <= relative_pos < 0.1 or 0.4 < relative_pos <= 0.5:
|
135 |
+
rewards.append(0.2)
|
136 |
+
elif 0.0 <= relative_pos < 0.05 or 0.5 < relative_pos <= 0.7:
|
137 |
+
rewards.append(0.1)
|
138 |
+
else:
|
139 |
+
rewards.append(0.0)
|
140 |
+
|
141 |
+
return rewards
|
142 |
+
|
143 |
+
def get_reward_funcs(script_args) -> list[Callable]:
|
144 |
+
"""Create a registry of available reward functions and return those specified in script_args."""
|
145 |
+
REWARD_FUNCS_REGISTRY = {
|
146 |
+
"accuracy": accuracy_reward,
|
147 |
+
"format": format_reward,
|
148 |
+
"code_execution": code_execution_reward,
|
149 |
+
"length": len_reward,
|
150 |
+
"code_ratio": code_ratio_reward,
|
151 |
+
"code_timing": code_timing_reward,
|
152 |
+
}
|
153 |
+
|
154 |
+
# Get the specified reward functions
|
155 |
+
reward_funcs = [REWARD_FUNCS_REGISTRY[func] for func in script_args.reward_funcs]
|
156 |
+
return reward_funcs
|