Upload templates.py with huggingface_hub
Browse files- templates.py +150 -33
templates.py
CHANGED
@@ -4,11 +4,13 @@ from dataclasses import field
|
|
4 |
from typing import Any, Dict, List, Optional, Union
|
5 |
|
6 |
from .artifact import Artifact
|
|
|
7 |
from .dataclass import NonPositionalField
|
8 |
from .instructions import Instruction, TextualInstruction
|
9 |
-
from .operator import
|
10 |
-
from .random_utils import
|
11 |
from .text_utils import split_words
|
|
|
12 |
|
13 |
|
14 |
class Renderer(ABC):
|
@@ -39,10 +41,14 @@ class RenderFormatTemplate(Renderer, StreamInstanceOperator):
|
|
39 |
random_reference: bool = False
|
40 |
|
41 |
def verify(self):
|
42 |
-
assert isinstance(
|
|
|
|
|
43 |
assert self.template is not None, "Template must be specified"
|
44 |
|
45 |
-
def process(
|
|
|
|
|
46 |
return self.render(instance)
|
47 |
|
48 |
def render(self, instance: Dict[str, Any]) -> Dict[str, Any]:
|
@@ -55,7 +61,7 @@ class RenderFormatTemplate(Renderer, StreamInstanceOperator):
|
|
55 |
if self.template.is_multi_reference:
|
56 |
references = targets
|
57 |
if self.random_reference:
|
58 |
-
target =
|
59 |
else:
|
60 |
if len(references) == 0:
|
61 |
raise ValueError("No references found")
|
@@ -87,7 +93,7 @@ class RenderAutoFormatTemplate(RenderFormatTemplate):
|
|
87 |
except:
|
88 |
pass
|
89 |
|
90 |
-
inputs =
|
91 |
|
92 |
return super().render({**instance, "inputs": inputs})
|
93 |
|
@@ -118,7 +124,12 @@ class RenderTemplatedICL(RenderAutoFormatTemplate):
|
|
118 |
|
119 |
example = super().render(instance)
|
120 |
|
121 |
-
input_str =
|
|
|
|
|
|
|
|
|
|
|
122 |
|
123 |
if self.instruction is not None:
|
124 |
source += self.instruction_prefix + self.instruction() + self.demo_separator
|
@@ -136,7 +147,9 @@ class RenderTemplatedICL(RenderAutoFormatTemplate):
|
|
136 |
)
|
137 |
|
138 |
if self.size_limiter is not None:
|
139 |
-
if not self.size_limiter.check(
|
|
|
|
|
140 |
continue
|
141 |
|
142 |
source += demo_str
|
@@ -155,7 +168,9 @@ class RenderTemplatedICL(RenderAutoFormatTemplate):
|
|
155 |
class InputOutputTemplate(Template):
|
156 |
input_format: str = None
|
157 |
output_format: str = None
|
158 |
-
postprocessors: List[str] = field(
|
|
|
|
|
159 |
|
160 |
def process_template(self, template: str, data: Dict[str, object]) -> str:
|
161 |
data = {k: ", ".join(v) if isinstance(v, list) else v for k, v in data.items()}
|
@@ -166,16 +181,91 @@ class InputOutputTemplate(Template):
|
|
166 |
return self.process_template(self.input_format, inputs)
|
167 |
except KeyError as e:
|
168 |
raise KeyError(
|
169 |
-
f"Available inputs are {inputs.keys()} but input format requires a different
|
170 |
-
)
|
171 |
|
172 |
def process_outputs(self, outputs: Dict[str, object]) -> str:
|
173 |
try:
|
174 |
return self.process_template(self.output_format, outputs)
|
175 |
except KeyError as e:
|
176 |
raise KeyError(
|
177 |
-
f"Available
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
178 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
179 |
|
180 |
def get_postprocessors(self) -> List[str]:
|
181 |
return self.postprocessors
|
@@ -188,10 +278,17 @@ class KeyValTemplate(Template):
|
|
188 |
outputs_key_val_seperator: str = ": "
|
189 |
use_keys_for_outputs: bool = False
|
190 |
|
191 |
-
postprocessors: List[str] = field(
|
|
|
|
|
192 |
|
193 |
-
def process_dict(
|
194 |
-
dic
|
|
|
|
|
|
|
|
|
|
|
195 |
pairs = []
|
196 |
for key, val in dic.items():
|
197 |
key_val = [key, val] if use_keys else [val]
|
@@ -221,9 +318,10 @@ class KeyValTemplate(Template):
|
|
221 |
class OutputQuantizingTemplate(InputOutputTemplate):
|
222 |
quantum: float = 0.1
|
223 |
|
224 |
-
def process_outputs(self, outputs: Dict[str, object]) ->
|
225 |
quantized_outputs = {
|
226 |
-
key: round(input_float / self.quantum) * self.quantum
|
|
|
227 |
}
|
228 |
return super().process_outputs(quantized_outputs)
|
229 |
|
@@ -235,12 +333,25 @@ class MultiLabelTemplate(InputOutputTemplate):
|
|
235 |
output_format = "{labels}"
|
236 |
empty_label = "None"
|
237 |
|
238 |
-
def process_outputs(self, outputs: Dict[str, object]) ->
|
239 |
labels = outputs[self.labels_field]
|
240 |
if len(labels) == 0:
|
241 |
labels = [self.empty_label]
|
242 |
labels_str = self.labels_seprator.join(labels)
|
243 |
-
return super().process_outputs({
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
244 |
|
245 |
|
246 |
def escape_chars(s, chars_to_escape):
|
@@ -296,13 +407,16 @@ class SpanLabelingTemplate(SpanLabelingBaseTemplate):
|
|
296 |
|
297 |
|
298 |
class SpanLabelingJsonTemplate(SpanLabelingBaseTemplate):
|
299 |
-
postprocessors = [
|
|
|
|
|
|
|
300 |
|
301 |
def span_label_pairs_to_targets(self, span_label_pairs):
|
302 |
groups = {}
|
303 |
for span, label in span_label_pairs:
|
304 |
if label not in groups:
|
305 |
-
groups[label] =
|
306 |
groups[label].append(span)
|
307 |
if len(groups) > 0:
|
308 |
targets = [json.dumps(groups)]
|
@@ -315,7 +429,9 @@ class AutoInputOutputTemplate(InputOutputTemplate):
|
|
315 |
def infer_input_format(self, inputs):
|
316 |
input_format = ""
|
317 |
for key in inputs.keys():
|
318 |
-
name = " ".join(
|
|
|
|
|
319 |
input_format += name + ": " + "{" + key + "}" + "\n"
|
320 |
self.input_format = input_format
|
321 |
|
@@ -332,21 +448,20 @@ class AutoInputOutputTemplate(InputOutputTemplate):
|
|
332 |
return self.input_format is not None and self.output_format is not None
|
333 |
|
334 |
|
335 |
-
from .collections import ListCollection
|
336 |
-
|
337 |
-
|
338 |
class TemplatesList(ListCollection):
|
339 |
def verify(self):
|
340 |
for template in self.items:
|
341 |
assert isinstance(template, Template)
|
342 |
|
343 |
|
344 |
-
def outputs_inputs2templates(
|
345 |
-
|
346 |
-
|
|
|
|
|
347 |
:param inputs: list of input formats (or one)
|
348 |
:param outputs: list of output formats (or one)
|
349 |
-
:return: TemplatesList of InputOutputTemplate
|
350 |
"""
|
351 |
templates = []
|
352 |
if isinstance(inputs, str):
|
@@ -367,8 +482,8 @@ def outputs_inputs2templates(inputs: Union[str, List], outputs: Union[str, List]
|
|
367 |
def instructions2templates(
|
368 |
instructions: List[TextualInstruction], templates: List[InputOutputTemplate]
|
369 |
) -> TemplatesList:
|
370 |
-
"""
|
371 |
-
|
372 |
:param instructions:
|
373 |
:param templates: strings containing {instuction} where the instruction should be placed
|
374 |
:return:
|
@@ -378,7 +493,9 @@ def instructions2templates(
|
|
378 |
for template in templates:
|
379 |
res_templates.append(
|
380 |
InputOutputTemplate(
|
381 |
-
input_format=template.input_format.replace(
|
|
|
|
|
382 |
output_format=template.output_format,
|
383 |
)
|
384 |
)
|
@@ -387,5 +504,5 @@ def instructions2templates(
|
|
387 |
|
388 |
class TemplatesDict(Dict):
|
389 |
def verify(self):
|
390 |
-
for
|
391 |
assert isinstance(template, Template)
|
|
|
4 |
from typing import Any, Dict, List, Optional, Union
|
5 |
|
6 |
from .artifact import Artifact
|
7 |
+
from .collections import ListCollection
|
8 |
from .dataclass import NonPositionalField
|
9 |
from .instructions import Instruction, TextualInstruction
|
10 |
+
from .operator import StreamInstanceOperator
|
11 |
+
from .random_utils import get_random
|
12 |
from .text_utils import split_words
|
13 |
+
from .type_utils import isoftype
|
14 |
|
15 |
|
16 |
class Renderer(ABC):
|
|
|
41 |
random_reference: bool = False
|
42 |
|
43 |
def verify(self):
|
44 |
+
assert isinstance(
|
45 |
+
self.template, Template
|
46 |
+
), "Template must be an instance of Template"
|
47 |
assert self.template is not None, "Template must be specified"
|
48 |
|
49 |
+
def process(
|
50 |
+
self, instance: Dict[str, Any], stream_name: Optional[str] = None
|
51 |
+
) -> Dict[str, Any]:
|
52 |
return self.render(instance)
|
53 |
|
54 |
def render(self, instance: Dict[str, Any]) -> Dict[str, Any]:
|
|
|
61 |
if self.template.is_multi_reference:
|
62 |
references = targets
|
63 |
if self.random_reference:
|
64 |
+
target = get_random().choice(references)
|
65 |
else:
|
66 |
if len(references) == 0:
|
67 |
raise ValueError("No references found")
|
|
|
93 |
except:
|
94 |
pass
|
95 |
|
96 |
+
inputs = dict(instance["inputs"].items())
|
97 |
|
98 |
return super().render({**instance, "inputs": inputs})
|
99 |
|
|
|
124 |
|
125 |
example = super().render(instance)
|
126 |
|
127 |
+
input_str = (
|
128 |
+
self.input_prefix
|
129 |
+
+ example["source"]
|
130 |
+
+ self.input_output_separator
|
131 |
+
+ self.output_prefix
|
132 |
+
)
|
133 |
|
134 |
if self.instruction is not None:
|
135 |
source += self.instruction_prefix + self.instruction() + self.demo_separator
|
|
|
147 |
)
|
148 |
|
149 |
if self.size_limiter is not None:
|
150 |
+
if not self.size_limiter.check(
|
151 |
+
source + demo_str + input_str + example["target"]
|
152 |
+
):
|
153 |
continue
|
154 |
|
155 |
source += demo_str
|
|
|
168 |
class InputOutputTemplate(Template):
|
169 |
input_format: str = None
|
170 |
output_format: str = None
|
171 |
+
postprocessors: List[str] = field(
|
172 |
+
default_factory=lambda: ["processors.to_string_stripped"]
|
173 |
+
)
|
174 |
|
175 |
def process_template(self, template: str, data: Dict[str, object]) -> str:
|
176 |
data = {k: ", ".join(v) if isinstance(v, list) else v for k, v in data.items()}
|
|
|
181 |
return self.process_template(self.input_format, inputs)
|
182 |
except KeyError as e:
|
183 |
raise KeyError(
|
184 |
+
f"Available inputs are {list(inputs.keys())} but input format requires a different ones: '{self.input_format}'"
|
185 |
+
) from e
|
186 |
|
187 |
def process_outputs(self, outputs: Dict[str, object]) -> str:
|
188 |
try:
|
189 |
return self.process_template(self.output_format, outputs)
|
190 |
except KeyError as e:
|
191 |
raise KeyError(
|
192 |
+
f"Available outputs are {outputs.keys()} but output format requires a different one: {self.output_format}"
|
193 |
+
) from e
|
194 |
+
|
195 |
+
def get_postprocessors(self) -> List[str]:
|
196 |
+
return self.postprocessors
|
197 |
+
|
198 |
+
|
199 |
+
class YesNoTemplate(Template):
|
200 |
+
"""A template for generating binary Yes/No questions asking whether an input text is of a specific class.
|
201 |
+
|
202 |
+
input_format:
|
203 |
+
Defines the format of the question.
|
204 |
+
class_field:
|
205 |
+
Defines the field that contains the name of the class that this template
|
206 |
+
asks of.
|
207 |
+
label_field:
|
208 |
+
Defines the field which contains the true label of the input text. If a gold label is equal to the
|
209 |
+
value in class_name, then the correct output is self.yes_answer (by default, "Yes").
|
210 |
+
Otherwise the correct output is self.no_answer (by default, "No").
|
211 |
+
yes_answer:
|
212 |
+
The output value for when the gold label equals self.class_name.
|
213 |
+
Defaults to "Yes".
|
214 |
+
no_answer:
|
215 |
+
The output value for when the gold label differs from self.class_name.
|
216 |
+
Defaults to "No".
|
217 |
+
"""
|
218 |
+
|
219 |
+
input_format: str = None
|
220 |
+
class_field: str = None
|
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 process_inputs(self, inputs: Dict[str, object]) -> str:
|
229 |
+
try:
|
230 |
+
data = {
|
231 |
+
k: ", ".join(v) if isinstance(v, list) else v for k, v in inputs.items()
|
232 |
+
}
|
233 |
+
return self.input_format.format(**data)
|
234 |
+
except KeyError as e:
|
235 |
+
raise RuntimeError(
|
236 |
+
f"Available inputs are {list(inputs.keys())} but input format requires a different one: {self.input_format}"
|
237 |
+
) from e
|
238 |
+
|
239 |
+
def process_outputs(self, outputs: Dict[str, object]) -> str:
|
240 |
+
try:
|
241 |
+
gold_class_names = outputs[self.label_field]
|
242 |
+
except KeyError as e:
|
243 |
+
raise RuntimeError(
|
244 |
+
f"Available outputs are {list(outputs.keys())}, missing required label field: '{self.label_field}'."
|
245 |
+
) from e
|
246 |
+
if not isinstance(gold_class_names, list) or not gold_class_names:
|
247 |
+
raise RuntimeError(
|
248 |
+
f"Unexpected value for gold_class_names: '{gold_class_names}'. Expected a non-empty list."
|
249 |
)
|
250 |
+
try:
|
251 |
+
queried_class_names = outputs[self.class_field]
|
252 |
+
except KeyError as e:
|
253 |
+
raise RuntimeError(
|
254 |
+
f"Available outputs are {list(outputs.keys())}, missing required class field: '{self.class_field}'."
|
255 |
+
) from e
|
256 |
+
if (
|
257 |
+
not queried_class_names
|
258 |
+
or not isinstance(queried_class_names, list)
|
259 |
+
or not len(queried_class_names) == 1
|
260 |
+
):
|
261 |
+
raise RuntimeError(
|
262 |
+
f"Unexpected value for queried_class_names: '{queried_class_names}'. Expected a list with one item."
|
263 |
+
)
|
264 |
+
queried_class_name = queried_class_names[0]
|
265 |
+
if queried_class_name in gold_class_names:
|
266 |
+
return self.yes_answer
|
267 |
+
|
268 |
+
return self.no_answer
|
269 |
|
270 |
def get_postprocessors(self) -> List[str]:
|
271 |
return self.postprocessors
|
|
|
278 |
outputs_key_val_seperator: str = ": "
|
279 |
use_keys_for_outputs: bool = False
|
280 |
|
281 |
+
postprocessors: List[str] = field(
|
282 |
+
default_factory=lambda: ["processors.to_string_stripped"]
|
283 |
+
)
|
284 |
|
285 |
+
def process_dict(
|
286 |
+
self, dic: Dict[str, object], key_val_sep, pairs_sep, use_keys
|
287 |
+
) -> str:
|
288 |
+
dic = {
|
289 |
+
k: ", ".join([str(vi) for vi in v]) if isinstance(v, list) else v
|
290 |
+
for k, v in dic.items()
|
291 |
+
}
|
292 |
pairs = []
|
293 |
for key, val in dic.items():
|
294 |
key_val = [key, val] if use_keys else [val]
|
|
|
318 |
class OutputQuantizingTemplate(InputOutputTemplate):
|
319 |
quantum: float = 0.1
|
320 |
|
321 |
+
def process_outputs(self, outputs: Dict[str, object]) -> str:
|
322 |
quantized_outputs = {
|
323 |
+
key: round(input_float / self.quantum) * self.quantum
|
324 |
+
for key, input_float in outputs.items()
|
325 |
}
|
326 |
return super().process_outputs(quantized_outputs)
|
327 |
|
|
|
333 |
output_format = "{labels}"
|
334 |
empty_label = "None"
|
335 |
|
336 |
+
def process_outputs(self, outputs: Dict[str, object]) -> str:
|
337 |
labels = outputs[self.labels_field]
|
338 |
if len(labels) == 0:
|
339 |
labels = [self.empty_label]
|
340 |
labels_str = self.labels_seprator.join(labels)
|
341 |
+
return super().process_outputs({self.labels_field: labels_str})
|
342 |
+
|
343 |
+
|
344 |
+
class MultiReferenceTemplate(InputOutputTemplate):
|
345 |
+
references_field: str = "references"
|
346 |
+
is_multi_reference = True
|
347 |
+
|
348 |
+
def process_outputs(self, outputs: Dict[str, object]) -> List[str]:
|
349 |
+
references = outputs[self.references_field]
|
350 |
+
if not isoftype(references, List[str]):
|
351 |
+
raise ValueError(
|
352 |
+
f"MultiReferenceTemplate requires that references field {self.references_field} is of type List[str]."
|
353 |
+
)
|
354 |
+
return references
|
355 |
|
356 |
|
357 |
def escape_chars(s, chars_to_escape):
|
|
|
407 |
|
408 |
|
409 |
class SpanLabelingJsonTemplate(SpanLabelingBaseTemplate):
|
410 |
+
postprocessors = [
|
411 |
+
"processors.load_json",
|
412 |
+
"processors.dict_of_lists_to_value_key_pairs",
|
413 |
+
]
|
414 |
|
415 |
def span_label_pairs_to_targets(self, span_label_pairs):
|
416 |
groups = {}
|
417 |
for span, label in span_label_pairs:
|
418 |
if label not in groups:
|
419 |
+
groups[label] = []
|
420 |
groups[label].append(span)
|
421 |
if len(groups) > 0:
|
422 |
targets = [json.dumps(groups)]
|
|
|
429 |
def infer_input_format(self, inputs):
|
430 |
input_format = ""
|
431 |
for key in inputs.keys():
|
432 |
+
name = " ".join(
|
433 |
+
word.lower().capitalize() for word in split_words(key) if word != " "
|
434 |
+
)
|
435 |
input_format += name + ": " + "{" + key + "}" + "\n"
|
436 |
self.input_format = input_format
|
437 |
|
|
|
448 |
return self.input_format is not None and self.output_format is not None
|
449 |
|
450 |
|
|
|
|
|
|
|
451 |
class TemplatesList(ListCollection):
|
452 |
def verify(self):
|
453 |
for template in self.items:
|
454 |
assert isinstance(template, Template)
|
455 |
|
456 |
|
457 |
+
def outputs_inputs2templates(
|
458 |
+
inputs: Union[str, List], outputs: Union[str, List]
|
459 |
+
) -> TemplatesList:
|
460 |
+
"""Combines input and output formats into their dot product.
|
461 |
+
|
462 |
:param inputs: list of input formats (or one)
|
463 |
:param outputs: list of output formats (or one)
|
464 |
+
:return: TemplatesList of InputOutputTemplate.
|
465 |
"""
|
466 |
templates = []
|
467 |
if isinstance(inputs, str):
|
|
|
482 |
def instructions2templates(
|
483 |
instructions: List[TextualInstruction], templates: List[InputOutputTemplate]
|
484 |
) -> TemplatesList:
|
485 |
+
"""Insert instructions into per demonstration templates.
|
486 |
+
|
487 |
:param instructions:
|
488 |
:param templates: strings containing {instuction} where the instruction should be placed
|
489 |
:return:
|
|
|
493 |
for template in templates:
|
494 |
res_templates.append(
|
495 |
InputOutputTemplate(
|
496 |
+
input_format=template.input_format.replace(
|
497 |
+
"{instruction}", instruction.text
|
498 |
+
),
|
499 |
output_format=template.output_format,
|
500 |
)
|
501 |
)
|
|
|
504 |
|
505 |
class TemplatesDict(Dict):
|
506 |
def verify(self):
|
507 |
+
for _key, template in self.items():
|
508 |
assert isinstance(template, Template)
|