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
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189 | @DataConverterRegistry.register("helpsteer2_pairwise")
class HelpSteer2PairwiseConverter(DataConverter):
"""
Converter for HelpSteer2 pairwise data format
Can handle data from both local files and HuggingFace Hub
Converts each data entry into two DataSamples with swapped responses
"""
def convert_to_data_sample(
self, data_dict: Dict[str, Any], source_info: Dict[str, Any]
) -> Union[DataSample, List[DataSample]]:
"""Convert HelpSteer2 pairwise data to DataSample format"""
try:
# Create input from prompt
data_input = [ChatMessage(role="user", content=data_dict["prompt"])]
# Determine preference based on preference_strength
preference_strength = data_dict.get("preference_strength", 0)
if preference_strength > 0:
# response_2 is better
preferred_in_original = "response_2"
elif preference_strength < 0:
# response_1 is better
preferred_in_original = "response_1"
else:
# tie
preferred_in_original = "tie"
data_samples = []
# Create first sample: response_A = response_1, response_B = response_2
sample1_id = hashlib.md5(f"{str(data_dict)}_sample1".encode()).hexdigest()
# Determine preferred for first sample
if preferred_in_original == "response_1":
preferred_1 = "A" # response_A (response_1) is preferred
elif preferred_in_original == "response_2":
preferred_1 = "B" # response_B (response_2) is preferred
else:
preferred_1 = "tie"
# Create outputs for first sample
output_1 = [
DataOutput(
answer=Step(
role="assistant",
content=data_dict["response_1"],
label={"response_type": "A"},
)
),
DataOutput(
answer=Step(
role="assistant",
content=data_dict["response_2"],
label={"response_type": "B"},
)
),
]
# Build metadata for first sample
metadata_1 = {
"raw_data": data_dict,
"load_strategy": "HelpSteer2PairwiseConverter",
"response_A": data_dict["response_1"],
"response_B": data_dict["response_2"],
"preferred": preferred_1,
"preference_strength": preference_strength,
"preference_statement": data_dict.get("preference_statement"),
"preference_elaboration": data_dict.get("preference_elaboration"),
"sample_type": "original_order",
}
# Add source-specific metadata
if source_info.get("load_type") == "local":
metadata_1.update(
{
"source_file_path": source_info.get("source_file_path"),
"load_type": "local",
}
)
elif source_info.get("load_type") == "huggingface":
metadata_1.update(
{
"dataset_name": source_info.get(
"dataset_name", "nvidia/HelpSteer2"
),
"dataset_config": source_info.get("dataset_config"),
"split": source_info.get("split", "train"),
"load_type": "huggingface",
}
)
sample_1 = DataSample(
unique_id=sample1_id,
input=data_input,
output=output_1,
source="helpsteer2_pairwise",
task_category="chat_pairwise",
metadata=metadata_1,
)
data_samples.append(sample_1)
# Create second sample: response_A = response_2, response_B = response_1 (swapped)
sample2_id = hashlib.md5(f"{str(data_dict)}_sample2".encode()).hexdigest()
# Determine preferred for second sample (swapped)
if preferred_in_original == "response_1":
preferred_2 = "B" # response_B (response_1) is preferred
elif preferred_in_original == "response_2":
preferred_2 = "A" # response_A (response_2) is preferred
else:
preferred_2 = "tie"
# Create outputs for second sample (swapped)
output_2 = [
DataOutput(
answer=Step(
role="assistant",
content=data_dict["response_2"],
label={"response_type": "A"},
)
),
DataOutput(
answer=Step(
role="assistant",
content=data_dict["response_1"],
label={"response_type": "B"},
)
),
]
# Build metadata for second sample
metadata_2 = {
"raw_data": data_dict,
"load_strategy": "HelpSteer2PairwiseConverter",
"response_A": data_dict["response_2"],
"response_B": data_dict["response_1"],
"preferred": preferred_2,
"preference_strength": preference_strength,
"preference_statement": data_dict.get("preference_statement"),
"preference_elaboration": data_dict.get("preference_elaboration"),
"sample_type": "swapped_order",
}
# Add source-specific metadata
if source_info.get("load_type") == "local":
metadata_2.update(
{
"source_file_path": source_info.get("source_file_path"),
"load_type": "local",
}
)
elif source_info.get("load_type") == "huggingface":
metadata_2.update(
{
"dataset_name": source_info.get(
"dataset_name", "nvidia/HelpSteer2"
),
"dataset_config": source_info.get("dataset_config"),
"split": source_info.get("split", "train"),
"load_type": "huggingface",
}
)
sample_2 = DataSample(
unique_id=sample2_id,
input=data_input,
output=output_2,
source="helpsteer2_pairwise",
task_category="chat_pairwise",
metadata=metadata_2,
)
data_samples.append(sample_2)
return data_samples
except Exception as e:
logger.error(f"Error creating HelpSteer2 Pairwise DataSample: {str(e)}")
return None
|