10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159 | @DataConverterRegistry.register("rmbbenchmark_bestofn")
class RMBBenchmarkBestOfNConverter(DataConverter):
"""
Unified converter for conversation data with conversation_input, bon_best and loser_list responses
"""
def convert_to_data_sample(
self, data_dict: Dict[str, Any], source_info: Dict[str, Any]
) -> DataSample:
"""Convert conversation data to DataSample format"""
# Generate unique id using bon_uid
if "bon_uid" in data_dict:
unique_id = str(data_dict["bon_uid"])
else:
# Use conversation_input content for generating hash
conversation_input = data_dict.get("conversation_input", [])
if (
conversation_input
and isinstance(conversation_input, list)
and len(conversation_input) > 0
):
content = str(conversation_input[0].get("content", ""))
else:
content = ""
unique_id = hashlib.md5(content.encode()).hexdigest()
# Create input from conversation_input
data_input = self._create_conversation_input(data_dict)
# Create outputs from bon_best and loser_list
data_output = self._create_conversation_output(data_dict)
try:
# Build metadata based on source type
metadata = {
"raw_data": data_dict,
"load_strategy": "RMBBenchmarkBestOfNConverter",
"category_path": data_dict.get("category_path"),
"bon_uid": data_dict.get("bon_uid"),
"bon_best_model": data_dict.get("bon_best", {}).get("llm_name")
if data_dict.get("bon_best")
else None,
"loser_models": [
item.get("llm_name")
for item in data_dict.get("loser_list", [])
if isinstance(item, dict)
],
"num_losers": len(data_dict.get("loser_list", [])),
}
# Add source-specific metadata
if source_info.get("load_type") == "local":
metadata.update(
{
"source_file_path": source_info.get("source_file_path"),
"load_type": "local",
}
)
elif source_info.get("load_type") == "huggingface":
metadata.update(
{
"dataset_name": source_info.get("dataset_name"),
"dataset_config": source_info.get("dataset_config"),
"split": source_info.get("split", "train"),
"load_type": "huggingface",
}
)
data_sample = DataSample(
unique_id=unique_id,
input=data_input,
output=data_output,
source="rewardbench",
task_category="conversation",
metadata=metadata,
)
return data_sample
except Exception as e:
logger.error(f"Error creating conversation DataSample: {str(e)}")
return None
def _create_conversation_input(
self, data_dict: Dict[str, Any]
) -> list[ChatMessage]:
"""Create DataInput from conversation_input"""
conversation_input = data_dict.get("conversation_input", [])
if isinstance(conversation_input, list):
history = []
for message in conversation_input:
if isinstance(message, dict):
role = message.get("role", "user")
content = message.get("content", "")
history.append(ChatMessage(role=role, content=content))
else:
history.append(ChatMessage(role="user", content=str(message)))
return history
else:
return [ChatMessage(role="user", content=str(conversation_input))]
def _create_conversation_output(
self, data_dict: Dict[str, Any]
) -> list[DataOutput]:
"""Create DataOutput list from bon_best and loser_list"""
outputs = []
# Handle bon_best
if "bon_best" in data_dict:
bon_best = data_dict["bon_best"]
if isinstance(bon_best, dict):
answer_content = bon_best.get("answer", "")
llm_name = bon_best.get("llm_name", "unknown")
outputs.append(
DataOutput(
answer=Step(
role="assistant",
content=str(answer_content),
label={
"preference": "chosen",
"model": llm_name,
"type": "bon_best",
},
),
)
)
# Handle loser_list
if "loser_list" in data_dict:
loser_list = data_dict["loser_list"]
if isinstance(loser_list, list):
for loser in loser_list:
if isinstance(loser, dict):
answer_content = loser.get("answer", "")
llm_name = loser.get("llm_name", "unknown")
outputs.append(
DataOutput(
answer=Step(
role="assistant",
content=str(answer_content),
label={
"preference": "rejected",
"model": llm_name,
"type": "loser",
},
),
)
)
return outputs
|