Upload templates.py with huggingface_hub
Browse files- templates.py +69 -42
templates.py
CHANGED
@@ -1,6 +1,5 @@
|
|
1 |
import json
|
2 |
from abc import abstractmethod
|
3 |
-
from dataclasses import field
|
4 |
from typing import Any, Dict, List, Optional, Tuple
|
5 |
|
6 |
from .collections import ListCollection
|
@@ -14,12 +13,21 @@ class Template(StreamInstanceOperator):
|
|
14 |
"""The role of template is to take the fields of every instance and verbalize it.
|
15 |
|
16 |
Meaning the template is taking the instance and generating source, target and references.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
"""
|
18 |
|
19 |
skip_rendered_instance: bool = NonPositionalField(default=True)
|
20 |
postprocessors: List[str] = NonPositionalField(
|
21 |
default_factory=lambda: ["processors.to_string_stripped"]
|
22 |
)
|
|
|
|
|
23 |
|
24 |
def process(
|
25 |
self, instance: Dict[str, Any], stream_name: Optional[str] = None
|
@@ -35,7 +43,7 @@ class Template(StreamInstanceOperator):
|
|
35 |
inputs = instance.get("inputs")
|
36 |
outputs = instance.get("outputs")
|
37 |
|
38 |
-
source = self.inputs_to_source(inputs)
|
39 |
target, references = self.outputs_to_target_and_references(outputs)
|
40 |
|
41 |
return {
|
@@ -43,10 +51,12 @@ class Template(StreamInstanceOperator):
|
|
43 |
"source": source,
|
44 |
"target": target,
|
45 |
"references": references,
|
|
|
|
|
46 |
}
|
47 |
|
48 |
@abstractmethod
|
49 |
-
def inputs_to_source(self, inputs: Dict[str, object]) -> str:
|
50 |
pass
|
51 |
|
52 |
@abstractmethod
|
@@ -72,13 +82,17 @@ class InputOutputTemplate(Template):
|
|
72 |
data = {k: ", ".join(v) if isinstance(v, list) else v for k, v in data.items()}
|
73 |
return template.format(**data)
|
74 |
|
75 |
-
def inputs_to_source(self, inputs: Dict[str, object]) -> str:
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
|
|
|
|
|
|
|
|
82 |
|
83 |
def outputs_to_target_and_references(self, outputs: Dict[str, object]) -> str:
|
84 |
try:
|
@@ -92,6 +106,25 @@ class InputOutputTemplate(Template):
|
|
92 |
return target, references
|
93 |
|
94 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
class MultipleChoiceTemplate(Template):
|
96 |
"""Formats the input (that specifies the question), the multiple choices to select the answer from, and specifies the field with the correct answer."""
|
97 |
|
@@ -149,19 +182,22 @@ class MultipleChoiceTemplate(Template):
|
|
149 |
)
|
150 |
return enumrated_choices
|
151 |
|
152 |
-
def inputs_to_source(self, inputs: Dict[str, object]) -> str:
|
153 |
choices = self.get_choices(inputs, self.source_choice_format)
|
154 |
inputs = {
|
155 |
"numerals": ",".join(self.get_choices(inputs, "{choice_numeral}")),
|
156 |
**inputs,
|
157 |
self.choices_field: self.choices_seperator.join(choices),
|
158 |
}
|
159 |
-
|
160 |
-
|
161 |
-
|
162 |
-
|
163 |
-
|
164 |
-
|
|
|
|
|
|
|
165 |
|
166 |
def outputs_to_target_and_references(self, outputs: Dict[str, object]) -> str:
|
167 |
target = outputs[self.target_field]
|
@@ -221,20 +257,20 @@ class YesNoTemplate(Template):
|
|
221 |
label_field: str = None
|
222 |
yes_answer: str = "Yes"
|
223 |
no_answer: str = "No"
|
224 |
-
postprocessors: List[str] = field(
|
225 |
-
default_factory=lambda: ["processors.to_string_stripped"]
|
226 |
-
)
|
227 |
|
228 |
-
def inputs_to_source(self, inputs: Dict[str, object]) -> str:
|
229 |
-
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
|
|
|
|
|
|
238 |
|
239 |
def outputs_to_target_and_references(self, outputs: Dict[str, object]) -> str:
|
240 |
try:
|
@@ -266,9 +302,6 @@ class YesNoTemplate(Template):
|
|
266 |
return self.yes_answer, [self.yes_answer]
|
267 |
return self.no_answer, [self.no_answer]
|
268 |
|
269 |
-
def get_postprocessors(self) -> List[str]:
|
270 |
-
return self.postprocessors
|
271 |
-
|
272 |
|
273 |
class KeyValTemplate(Template):
|
274 |
"""Generate field 'source' from fields designated as input, and fields 'target' and 'references' from fields designated as output, of the processed instance.
|
@@ -282,10 +315,6 @@ class KeyValTemplate(Template):
|
|
282 |
outputs_key_val_seperator: str = ": "
|
283 |
use_keys_for_outputs: bool = False
|
284 |
|
285 |
-
postprocessors: List[str] = field(
|
286 |
-
default_factory=lambda: ["processors.to_string_stripped"]
|
287 |
-
)
|
288 |
-
|
289 |
def process_dict(
|
290 |
self, dic: Dict[str, object], key_val_sep, pairs_sep, use_keys
|
291 |
) -> str:
|
@@ -299,13 +328,14 @@ class KeyValTemplate(Template):
|
|
299 |
pairs.append(key_val_sep.join(key_val))
|
300 |
return pairs_sep.join(pairs)
|
301 |
|
302 |
-
def inputs_to_source(self, inputs: Dict[str, object]) -> str:
|
303 |
-
|
304 |
inputs,
|
305 |
key_val_sep=self.key_val_seperator,
|
306 |
pairs_sep=self.pairs_seperator,
|
307 |
use_keys=self.use_keys_for_inputs,
|
308 |
)
|
|
|
309 |
|
310 |
def outputs_to_target_and_references(self, outputs: Dict[str, object]) -> str:
|
311 |
target = self.process_dict(
|
@@ -316,9 +346,6 @@ class KeyValTemplate(Template):
|
|
316 |
)
|
317 |
return target, [target]
|
318 |
|
319 |
-
def get_postprocessors(self) -> List[str]:
|
320 |
-
return self.postprocessors
|
321 |
-
|
322 |
|
323 |
class OutputQuantizingTemplate(InputOutputTemplate):
|
324 |
quantum: float = 0.1
|
|
|
1 |
import json
|
2 |
from abc import abstractmethod
|
|
|
3 |
from typing import Any, Dict, List, Optional, Tuple
|
4 |
|
5 |
from .collections import ListCollection
|
|
|
13 |
"""The role of template is to take the fields of every instance and verbalize it.
|
14 |
|
15 |
Meaning the template is taking the instance and generating source, target and references.
|
16 |
+
|
17 |
+
Args:
|
18 |
+
skip_rendered_instance (bool): if "source", "target", and "references" are already defined fields in the instance, skip its processing
|
19 |
+
postprocessors: a list of strings being artifact names of text processors, to be applied on the model output
|
20 |
+
instruction: a formatting string that yields an instruction with potential participation of values from the "inputs" part of the instance
|
21 |
+
target_prefix: a string to be used to format the prompt. Not a formatting string.
|
22 |
+
|
23 |
"""
|
24 |
|
25 |
skip_rendered_instance: bool = NonPositionalField(default=True)
|
26 |
postprocessors: List[str] = NonPositionalField(
|
27 |
default_factory=lambda: ["processors.to_string_stripped"]
|
28 |
)
|
29 |
+
instruction: str = NonPositionalField(default_factory=lambda: "")
|
30 |
+
target_prefix: str = NonPositionalField(default_factory=lambda: "")
|
31 |
|
32 |
def process(
|
33 |
self, instance: Dict[str, Any], stream_name: Optional[str] = None
|
|
|
43 |
inputs = instance.get("inputs")
|
44 |
outputs = instance.get("outputs")
|
45 |
|
46 |
+
source, instruction = self.inputs_to_source(inputs)
|
47 |
target, references = self.outputs_to_target_and_references(outputs)
|
48 |
|
49 |
return {
|
|
|
51 |
"source": source,
|
52 |
"target": target,
|
53 |
"references": references,
|
54 |
+
"instruction": instruction,
|
55 |
+
"target_prefix": self.target_prefix.format(**inputs),
|
56 |
}
|
57 |
|
58 |
@abstractmethod
|
59 |
+
def inputs_to_source(self, inputs: Dict[str, object]) -> Tuple[str, str]:
|
60 |
pass
|
61 |
|
62 |
@abstractmethod
|
|
|
82 |
data = {k: ", ".join(v) if isinstance(v, list) else v for k, v in data.items()}
|
83 |
return template.format(**data)
|
84 |
|
85 |
+
def inputs_to_source(self, inputs: Dict[str, object]) -> Tuple[str, str]:
|
86 |
+
formatted = []
|
87 |
+
for formatting in [self.input_format, self.instruction]:
|
88 |
+
try:
|
89 |
+
formatted.append(self.process_template(formatting, inputs))
|
90 |
+
except KeyError as e:
|
91 |
+
raise KeyError(
|
92 |
+
f"Available inputs are {list(inputs.keys())} but input format requires a different ones: '{formatting}'"
|
93 |
+
) from e
|
94 |
+
|
95 |
+
return tuple(formatted)
|
96 |
|
97 |
def outputs_to_target_and_references(self, outputs: Dict[str, object]) -> str:
|
98 |
try:
|
|
|
106 |
return target, references
|
107 |
|
108 |
|
109 |
+
class InputOutputReferenceTemplate(InputOutputTemplate):
|
110 |
+
reference: str
|
111 |
+
|
112 |
+
def outputs_to_target_and_references(self, outputs: Dict[str, object]) -> str:
|
113 |
+
output_fields = {}
|
114 |
+
for name, val in [
|
115 |
+
("target", self.output_format),
|
116 |
+
("reference", self.reference),
|
117 |
+
]:
|
118 |
+
try:
|
119 |
+
result = self.process_template(val, outputs)
|
120 |
+
output_fields[name] = result
|
121 |
+
except KeyError as e:
|
122 |
+
raise KeyError(
|
123 |
+
f"Available outputs are {outputs.keys()} but {name} requires a different one: {val}"
|
124 |
+
) from e
|
125 |
+
return output_fields["target"], [output_fields["reference"]]
|
126 |
+
|
127 |
+
|
128 |
class MultipleChoiceTemplate(Template):
|
129 |
"""Formats the input (that specifies the question), the multiple choices to select the answer from, and specifies the field with the correct answer."""
|
130 |
|
|
|
182 |
)
|
183 |
return enumrated_choices
|
184 |
|
185 |
+
def inputs_to_source(self, inputs: Dict[str, object]) -> Tuple[str, str]:
|
186 |
choices = self.get_choices(inputs, self.source_choice_format)
|
187 |
inputs = {
|
188 |
"numerals": ",".join(self.get_choices(inputs, "{choice_numeral}")),
|
189 |
**inputs,
|
190 |
self.choices_field: self.choices_seperator.join(choices),
|
191 |
}
|
192 |
+
formatted = []
|
193 |
+
for formatting in [self.input_format, self.instruction]:
|
194 |
+
try:
|
195 |
+
formatted.append(formatting.format(**inputs))
|
196 |
+
except KeyError as e:
|
197 |
+
raise KeyError(
|
198 |
+
f"Available inputs are {inputs.keys()} but input format requires a different one: {formatting}"
|
199 |
+
) from e
|
200 |
+
return tuple(formatted)
|
201 |
|
202 |
def outputs_to_target_and_references(self, outputs: Dict[str, object]) -> str:
|
203 |
target = outputs[self.target_field]
|
|
|
257 |
label_field: str = None
|
258 |
yes_answer: str = "Yes"
|
259 |
no_answer: str = "No"
|
|
|
|
|
|
|
260 |
|
261 |
+
def inputs_to_source(self, inputs: Dict[str, object]) -> Tuple[str, str]:
|
262 |
+
data = {
|
263 |
+
k: ", ".join(v) if isinstance(v, list) else v for k, v in inputs.items()
|
264 |
+
}
|
265 |
+
formatted = []
|
266 |
+
for formatting in [self.input_format, self.instruction]:
|
267 |
+
try:
|
268 |
+
formatted.append(formatting.format(**data))
|
269 |
+
except KeyError as e:
|
270 |
+
raise RuntimeError(
|
271 |
+
f"Available inputs are {list(inputs.keys())} but input format requires a different one: {formatting}"
|
272 |
+
) from e
|
273 |
+
return tuple(formatted)
|
274 |
|
275 |
def outputs_to_target_and_references(self, outputs: Dict[str, object]) -> str:
|
276 |
try:
|
|
|
302 |
return self.yes_answer, [self.yes_answer]
|
303 |
return self.no_answer, [self.no_answer]
|
304 |
|
|
|
|
|
|
|
305 |
|
306 |
class KeyValTemplate(Template):
|
307 |
"""Generate field 'source' from fields designated as input, and fields 'target' and 'references' from fields designated as output, of the processed instance.
|
|
|
315 |
outputs_key_val_seperator: str = ": "
|
316 |
use_keys_for_outputs: bool = False
|
317 |
|
|
|
|
|
|
|
|
|
318 |
def process_dict(
|
319 |
self, dic: Dict[str, object], key_val_sep, pairs_sep, use_keys
|
320 |
) -> str:
|
|
|
328 |
pairs.append(key_val_sep.join(key_val))
|
329 |
return pairs_sep.join(pairs)
|
330 |
|
331 |
+
def inputs_to_source(self, inputs: Dict[str, object]) -> Tuple[str, str]:
|
332 |
+
ret = self.process_dict(
|
333 |
inputs,
|
334 |
key_val_sep=self.key_val_seperator,
|
335 |
pairs_sep=self.pairs_seperator,
|
336 |
use_keys=self.use_keys_for_inputs,
|
337 |
)
|
338 |
+
return (ret, ret)
|
339 |
|
340 |
def outputs_to_target_and_references(self, outputs: Dict[str, object]) -> str:
|
341 |
target = self.process_dict(
|
|
|
346 |
)
|
347 |
return target, [target]
|
348 |
|
|
|
|
|
|
|
349 |
|
350 |
class OutputQuantizingTemplate(InputOutputTemplate):
|
351 |
quantum: float = 0.1
|