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
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211 | @DataConverterRegistry.register("*")
class GenericConverter(DataConverter):
"""
Generic converter that automatically handles diverse HuggingFace dataset formats.
Acts as a fallback converter when no specific format converter is available.
Intelligently extracts input/output pairs from common field names and structures.
Supported Input Patterns:
- Fields: prompt, question, input, text, instruction (for input)
- Fields: response, answer, output, completion (for output)
- Messages: array of role/content objects for conversations
Output: DataSample with auto-detected task category and structured data
"""
def convert_to_data_sample(
self, data_dict: Dict[str, Any], source_info: Dict[str, Any]
) -> DataSample:
"""
Convert generic HuggingFace data dictionary to standardized DataSample format.
Automatically detects input/output patterns from common field names,
determines task category, and creates appropriate data structure.
Args:
data_dict: Raw data dictionary from HuggingFace dataset
source_info: Source metadata including dataset name, config, split info
Returns:
DataSample with auto-detected structure and task category
Returns None if input/output extraction fails
"""
# Generate unique id
content = str(data_dict)
unique_id = hashlib.md5(content.encode()).hexdigest()
try:
# Try to extract input from common field names
input_data = self._extract_input(data_dict)
if not input_data:
logger.warning(f"Could not extract input from data: {data_dict}")
return None
# Try to extract output from common field names
output_data = self._extract_output(data_dict)
if not output_data:
logger.warning(f"Could not extract output from data: {data_dict}")
return None
# Determine task category
task_category = self._determine_task_category(data_dict)
# Build metadata based on source type
metadata = {
"raw_data": data_dict,
"load_strategy": "GenericConverter",
"task_category": task_category,
}
# 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=input_data,
output=output_data,
source=source_info.get("dataset_name", "generic"),
task_category=task_category,
metadata=metadata,
)
return data_sample
except Exception as e:
logger.error(f"Error creating generic DataSample: {str(e)}")
return None
def _extract_input(self, data_dict: Dict[str, Any]) -> list[ChatMessage]:
"""
Extract input messages from data using common field name patterns.
Searches for standard input field names and converts to ChatMessage format.
Handles both single-field inputs and conversation message arrays.
Args:
data_dict: Raw data dictionary to extract input from
Returns:
List of ChatMessage objects representing the input context
"""
input_data = []
# Common input field names
for field in ["prompt", "question", "input", "text", "instruction"]:
if field in data_dict and data_dict[field]:
input_data.append(
ChatMessage(role="user", content=str(data_dict[field]))
)
break
# Handle conversation/messages format
if "messages" in data_dict:
messages = data_dict["messages"]
if isinstance(messages, list):
for msg in messages:
if isinstance(msg, dict):
role = msg.get("role", "user")
content = msg.get("content", str(msg))
if role in ["user", "system"]: # Only include input messages
input_data.append(ChatMessage(role=role, content=content))
return input_data
def _extract_output(self, data_dict: Dict[str, Any]) -> list[DataOutput]:
"""
Extract output responses from data using common field name patterns.
Searches for standard output field names and creates DataOutput objects
with Step components for response evaluation.
Args:
data_dict: Raw data dictionary to extract output from
Returns:
List of DataOutput objects representing expected responses
"""
outputs = []
# Common output field names
for field in ["response", "answer", "output", "completion"]:
if field in data_dict and data_dict[field]:
outputs.append(
DataOutput(
answer=Step(role="assistant", content=str(data_dict[field]))
)
)
break
# Handle messages format for assistant responses
if "messages" in data_dict and not outputs:
messages = data_dict["messages"]
if isinstance(messages, list):
for msg in messages:
if isinstance(msg, dict) and msg.get("role") == "assistant":
outputs.append(
DataOutput(
answer=Step(
role="assistant",
content=str(msg.get("content", "")),
)
)
)
return outputs
def _determine_task_category(self, data_dict: Dict[str, Any]) -> str:
"""
Automatically determine task category from data field patterns.
Analyzes field names and structure to classify the type of task
for appropriate processing and evaluation strategies.
Args:
data_dict: Raw data dictionary to analyze
Returns:
String identifier for the detected task category
"""
# Check for explicit task category
if "task_category" in data_dict:
return str(data_dict["task_category"])
# Infer from field names
if any(field in data_dict for field in ["messages", "conversation"]):
return "chat"
elif any(field in data_dict for field in ["question", "answer"]):
return "qa"
elif any(field in data_dict for field in ["instruction", "completion"]):
return "instruction_following"
else:
return "general"
|