import json from abc import ABC, abstractmethod from dataclasses import field from typing import Any, Dict, List, Optional, Union from .artifact import Artifact from .dataclass import NonPositionalField from .instructions import Instruction, TextualInstruction from .operator import InstanceOperatorWithGlobalAccess, StreamInstanceOperator from .random_utils import random from .text_utils import split_words class Renderer(ABC): @abstractmethod def get_postprocessors(self) -> List[str]: pass class Template(Artifact): is_multi_target: bool = NonPositionalField(default=False) is_multi_reference: bool = NonPositionalField(default=False) @abstractmethod def process_inputs(self, inputs: Dict[str, object]) -> Dict[str, object]: pass @abstractmethod def process_outputs(self, outputs: Dict[str, object]) -> Dict[str, object]: pass @abstractmethod def get_postprocessors(self) -> List[str]: pass class RenderFormatTemplate(Renderer, StreamInstanceOperator): template: Template = None random_reference: bool = False def verify(self): assert isinstance(self.template, Template), "Template must be an instance of Template" assert self.template is not None, "Template must be specified" def process(self, instance: Dict[str, Any], stream_name: str = None) -> Dict[str, Any]: return self.render(instance) def render(self, instance: Dict[str, Any]) -> Dict[str, Any]: inputs = instance.pop("inputs") outputs = instance.pop("outputs") source = self.template.process_inputs(inputs) targets = self.template.process_outputs(outputs) if self.template.is_multi_reference: references = targets if self.random_reference: target = random.choice(references) else: if len(references) == 0: raise ValueError("No references found") target = references[0] else: references = [targets] target = targets return { **instance, "source": source, "target": target, "references": references, } def get_postprocessors(self) -> List[str]: return self.template.get_postprocessors() class RenderAutoFormatTemplate(RenderFormatTemplate): def prepare(self): if self.template is None: self.template = AutoInputOutputTemplate() def render(self, instance: Dict[str, object]) -> Dict[str, object]: try: if not self.template.is_complete(): self.template.infer_missing(instance["inputs"], instance["outputs"]) except: pass inputs = {key: value for key, value in instance["inputs"].items()} return super().render({**instance, "inputs": inputs}) class CharacterSizeLimiter(Artifact): limit: int = 1000 def check(self, text: str) -> bool: return len(text) <= self.limit class RenderTemplatedICL(RenderAutoFormatTemplate): instruction: Instruction = None input_prefix: str = "" output_prefix: str = "" target_prefix: str = " " instruction_prefix: str = "" demos_field: str = None size_limiter: Artifact = None input_output_separator: str = "\n" demo_separator: str = "\n\n" system_prompt: str = None def render(self, instance: Dict[str, object]) -> Dict[str, object]: demos = instance.pop(self.demos_field, []) source = "" example = super().render(instance) input_str = self.input_prefix + example["source"] + self.input_output_separator + self.output_prefix if self.instruction is not None: source += self.instruction_prefix + self.instruction() + self.demo_separator for demo_instance in demos: demo_example = super().render(demo_instance) demo_str = ( self.input_prefix + demo_example["source"] + self.input_output_separator + self.output_prefix + self.target_prefix + demo_example["target"] + self.demo_separator ) if self.size_limiter is not None: if not self.size_limiter.check(source + demo_str + input_str + example["target"]): continue source += demo_str source += input_str if self.system_prompt is not None: source = self.system_prompt.format(source) return { **example, "source": source, } class InputOutputTemplate(Template): input_format: str = None output_format: str = None postprocessors: List[str] = field(default_factory=lambda: ["processors.to_string_stripped"]) def process_template(self, template: str, data: Dict[str, object]) -> str: data = {k: ", ".join(v) if isinstance(v, list) else v for k, v in data.items()} return template.format(**data) def process_inputs(self, inputs: Dict[str, object]) -> str: try: return self.process_template(self.input_format, inputs) except KeyError as e: raise KeyError( f"Available inputs are {inputs.keys()} but input format requires a different one: {self.input_format}" ) def process_outputs(self, outputs: Dict[str, object]) -> str: try: return self.process_template(self.output_format, outputs) except KeyError as e: raise KeyError( f"Available inputs are {outputs.keys()} but output format requires a different one: {self.output_format}" ) def get_postprocessors(self) -> List[str]: return self.postprocessors class KeyValTemplate(Template): pairs_seperator: str = ", " key_val_seperator: str = ": " use_keys_for_inputs: bool = True outputs_key_val_seperator: str = ": " use_keys_for_outputs: bool = False postprocessors: List[str] = field(default_factory=lambda: ["processors.to_string_stripped"]) def process_dict(self, dic: Dict[str, object], key_val_sep, pairs_sep, use_keys) -> str: dic = {k: ", ".join(v) if isinstance(v, list) else v for k, v in dic.items()} pairs = [] for key, val in dic.items(): key_val = [key, val] if use_keys else [val] pairs.append(key_val_sep.join(key_val)) return pairs_sep.join(pairs) def process_inputs(self, inputs: Dict[str, object]) -> str: return self.process_dict( inputs, key_val_sep=self.key_val_seperator, pairs_sep=self.pairs_seperator, use_keys=self.use_keys_for_inputs, ) def process_outputs(self, outputs: Dict[str, object]) -> str: return self.process_dict( outputs, key_val_sep=self.key_val_seperator, pairs_sep=self.pairs_seperator, use_keys=self.use_keys_for_outputs, ) def get_postprocessors(self) -> List[str]: return self.postprocessors class OutputQuantizingTemplate(InputOutputTemplate): quantum: float = 0.1 def process_outputs(self, outputs: Dict[str, object]) -> Dict[str, object]: quantized_outputs = { key: round(input_float / self.quantum) * self.quantum for key, input_float in outputs.items() } return super().process_outputs(quantized_outputs) class MultiLabelTemplate(InputOutputTemplate): labels_field: str = "labels" labels_seprator: str = ", " postprocessors = ["processors.to_list_by_comma"] output_format = "{labels}" empty_label = "None" def process_outputs(self, outputs: Dict[str, object]) -> Dict[str, object]: labels = outputs[self.labels_field] if len(labels) == 0: labels = [self.empty_label] labels_str = self.labels_seprator.join(labels) return super().process_outputs({"labels": labels_str}) def escape_chars(s, chars_to_escape): for char in chars_to_escape: s = s.replace(char, f"\\{char}") return s class SpanLabelingBaseTemplate(MultiLabelTemplate): spans_starts_field: str = "spans_starts" spans_ends_field: str = "spans_ends" text_field: str = "text" labels_support: list = None def extract_span_label_pairs(self, outputs): spans_starts = outputs[self.spans_starts_field] spans_ends = outputs[self.spans_ends_field] text = outputs[self.text_field] labels = outputs[self.labels_field] spans = [] for span_start, span_end, label in zip(spans_starts, spans_ends, labels): if self.labels_support is None or label in self.labels_support: spans.append((span_start, span_end, text[span_start:span_end], label)) for span in sorted(spans): if self.labels_support is None or span[3] in self.labels_support: yield span[2], span[3] def process_outputs(self, outputs: Dict[str, object]) -> Dict[str, object]: span_lables_pairs = self.extract_span_label_pairs(outputs) targets = self.span_label_pairs_to_targets(span_lables_pairs) return super().process_outputs({"labels": targets}) @abstractmethod def span_label_pairs_to_targets(self, pairs): pass class SpanLabelingTemplate(SpanLabelingBaseTemplate): span_label_format: str = "{span}: {label}" escape_characters: List[str] = [":", ","] postprocessors = ["processors.to_span_label_pairs"] def span_label_pairs_to_targets(self, span_label_pairs): targets = [] for span, label in span_label_pairs: if self.escape_characters is not None: span = escape_chars(span, self.escape_characters) target = self.span_label_format.format(span=span, label=label) targets.append(target) return targets class SpanLabelingJsonTemplate(SpanLabelingBaseTemplate): postprocessors = ["processors.load_json", "processors.dict_of_lists_to_value_key_pairs"] def span_label_pairs_to_targets(self, span_label_pairs): groups = {} for span, label in span_label_pairs: if label not in groups: groups[label] = list() groups[label].append(span) if len(groups) > 0: targets = [json.dumps(groups)] else: targets = [] return targets class AutoInputOutputTemplate(InputOutputTemplate): def infer_input_format(self, inputs): input_format = "" for key in inputs.keys(): name = " ".join(word.lower().capitalize() for word in split_words(key) if word != " ") input_format += name + ": " + "{" + key + "}" + "\n" self.input_format = input_format def infer_output_format(self, outputs): self.output_format = "{" + next(iter(outputs.keys())) + "}" def infer_missing(self, inputs, outputs): if self.input_format is None: self.infer_input_format(inputs) if self.output_format is None: self.infer_output_format(outputs) def is_complete(self): return self.input_format is not None and self.output_format is not None from .collections import ListCollection class TemplatesList(ListCollection): def verify(self): for template in self.items: assert isinstance(template, Template) def outputs_inputs2templates(inputs: Union[str, List], outputs: Union[str, List]) -> TemplatesList: """ combines input and output formats into their dot product :param inputs: list of input formats (or one) :param outputs: list of output formats (or one) :return: TemplatesList of InputOutputTemplate """ templates = [] if isinstance(inputs, str): inputs = [inputs] if isinstance(outputs, str): outputs = [outputs] for input in inputs: for output in outputs: templates.append( InputOutputTemplate( input_format=input.strip(), output_format=output.strip(), ), ) return TemplatesList(templates) def instructions2templates( instructions: List[TextualInstruction], templates: List[InputOutputTemplate] ) -> TemplatesList: """ Insert instructions into per demonstration templates :param instructions: :param templates: strings containing {instuction} where the instruction should be placed :return: """ res_templates = [] for instruction in instructions: for template in templates: res_templates.append( InputOutputTemplate( input_format=template.input_format.replace("{instruction}", instruction.text), output_format=template.output_format, ) ) return TemplatesList(templates) class TemplatesDict(Dict): def verify(self): for key, template in self.items(): assert isinstance(template, Template)