diff --git a/scripts/test_knowledge.py b/scripts/test_knowledge.py index e800d3b2..d777c8c3 100644 --- a/scripts/test_knowledge.py +++ b/scripts/test_knowledge.py @@ -24,13 +24,13 @@ samples = [ { - "question_1": "what is the location of the tubal tonsils?", - "response_1": "The location of the tubal tonsils is the roof of the pharynx.", - "question_2": "How long does the adenoid grow?", + "icl_query_1": "what is the location of the tubal tonsils?", + "icl_response_1": "The location of the tubal tonsils is the roof of the pharynx.", + "icl_query_2": "How long does the adenoid grow?", "task_description": "Teaching about human anatomy, specifically tonsils", - "response_2": "The adenoid grows until the age of 5, starts to shrink at the age of 7 and becomes small in adulthood.", - "question_3": "What is the immune systems first line of defense against ingested or inhaled foreign pathogens?", - "response_3": "The tonsils are the immune systems first line of defense.", + "icl_response_2": "The adenoid grows until the age of 5, starts to shrink at the age of 7 and becomes small in adulthood.", + "icl_query_3": "What is the immune systems first line of defense against ingested or inhaled foreign pathogens?", + "icl_response_3": "The tonsils are the immune systems first line of defense.", "document": "The **tonsils** are a set of lymphoid organs facing into the aerodigestive tract, which is known as Waldeyer's tonsillar ring and consists of the adenoid tonsil or pharyngeal tonsil, two tubal tonsils, two palatine tonsils, and the lingual tonsils. These organs play an important role in the immune system. When used unqualified, the term most commonly refers specifically to the palatine tonsils, which are two lymphoid organs situated at either side of the back of the human throat. The palatine tonsils and the adenoid tonsil are organs consisting of lymphoepithelial tissue located near the oropharynx and nasopharynx parts of the throat", "domain": "textbook", } diff --git a/src/instructlab/sdg/configs/knowledge/generate_questions_responses.yaml b/src/instructlab/sdg/configs/knowledge/generate_questions_responses.yaml index b424f517..4d4a49ef 100644 --- a/src/instructlab/sdg/configs/knowledge/generate_questions_responses.yaml +++ b/src/instructlab/sdg/configs/knowledge/generate_questions_responses.yaml @@ -48,31 +48,33 @@ examples: | For this {domain} domain here are some sample questions: [Start of Question] - {question_1} + {icl_query_1} [End of Question] [Start of Response] - {response_1} + {icl_response_1} [End of Response] [Start of Question] - {question_2} + {icl_query_2} [End of Question] [Start of Response] - {response_2} + {icl_response_2} [End of Response] [Start of Question] - {question_3} + {icl_query_3} [End of Question] [Start of Response] - {response_3} + {icl_response_3} [End of Response] +generation: | + Now generate the question and answer pairs, remember to follow the principles mentioned above and use the same format as the examples. Remember to use the same style and format as the example above. + Here is the document: {document} -generation: | - Now generate the question and answer pairs, remember to follow the principles mentioned above and use the same format as the examples. Remember to use the same style and format as the example above. Return each question between [Start of Question] and [End of Question] tags and answer between [Start of Response] and [End of Response] tags. + Return each question between [Start of Question] and [End of Question] tags and answer between [Start of Response] and [End of Response] tags. start_tags: ["[Start of Question]", "[Start of Response]"] end_tags: ["[End of Question]", "[End of Response]"] diff --git a/src/instructlab/sdg/llmblock.py b/src/instructlab/sdg/llmblock.py index ad429b75..45a20343 100644 --- a/src/instructlab/sdg/llmblock.py +++ b/src/instructlab/sdg/llmblock.py @@ -1,5 +1,6 @@ # SPDX-License-Identifier: Apache-2.0 # Standard +from typing import Any, Dict import re # Third Party @@ -56,9 +57,8 @@ def _parse(self, generated_string) -> dict: pattern = re.escape(start_tag) + r"(.*?)" + re.escape(end_tag) all_matches = re.findall(pattern, generated_string, re.DOTALL) matches[output_col] = ( - [match.strip() for match in all_matches] if all_matches else None + [match.strip() for match in all_matches] if all_matches else [] ) - return matches def _generate(self, samples, **gen_kwargs) -> list: @@ -86,7 +86,7 @@ def generate(self, samples, **gen_kwargs) -> Dataset: if (num_samples is not None) and ("num_samples" not in samples.column_names): samples = samples.add_column("num_samples", [num_samples] * len(samples)) - # validate the each sample + # validate each sample for sample in samples: if not self._validate(self.prompt_template, sample): return None @@ -107,3 +107,91 @@ def generate(self, samples, **gen_kwargs) -> Dataset: new_data.append({**sample, **dict(zip(parsed_outputs.keys(), values))}) return Dataset.from_list(new_data) + + +class ConditionalLLMBlock(LLMBlock): + def __init__( + self, + block_name, + config_paths, + client, + model_id, + output_cols, + selector_column_name, + parser_name, + model_prompt="{prompt}", + **batch_kwargs, + ) -> None: + super().__init__( + block_name, + config_paths[0][0], + client, + model_id, + output_cols, + model_prompt=model_prompt, + **batch_kwargs, + ) + self.selector_column_name = selector_column_name + self.prompt_template = {} + self.parser_name = parser_name + if len(config_paths) == 1 and config_paths[0][1] == "All": + self.prompt_template = self.prompt_struct.format(**self.block_config) + else: + for config, config_key in config_paths: + self.prompt_template[config_key] = self.prompt_struct.format( + **self._load_config(config) + ) + + def _parse(self, generated_string): + if self.parser_name == "default": + return super()._parse(generated_string) + if self.parser_name == "multi-line-logical-section": + return { + self.output_cols[0]: self.extract_multiline_logical_section( + generated_string + ) + } + + def extract_multiline_logical_section(self, text): + """ + Extracts multi-line points from the provided text into a list, removing the point numbers. + + Args: + text (str): The input text containing multi-line points. + + Returns: + list: A list of multi-line points without the point numbers. + """ + pattern = re.compile( + r"## Logical Section \d+: (.*?)(?=## Logical Section \d+:|$)", re.DOTALL + ) + sections = pattern.findall(text) + + return sections + + def _generate(self, samples, **gen_kwargs) -> str: + if isinstance(self.prompt_template, dict): + prompts = [ + self.model_prompt.format( + prompt=self.prompt_template[sample[self.selector_column_name]] + .format(**sample) + .strip() + ) + for sample in samples + ] + else: + prompts = [ + self.model_prompt.format( + prompt=self.prompt_template.format(**sample).strip() + ) + for sample in samples + ] + response = self.client.completions.create( + prompt=prompts, **{**self.defaults, **gen_kwargs} + ) + return [choice.text.strip() for choice in response.choices] + + def _validate(self, prompt_template: str, input_dict: Dict[str, Any], extra_arg=None) -> bool: + if isinstance(prompt_template, dict): + prompt_template = prompt_template[input_dict[self.selector_column_name]] + return super()._validate(prompt_template, input_dict) diff --git a/src/instructlab/sdg/pipeline.py b/src/instructlab/sdg/pipeline.py index 0de65d1b..fc93f78d 100644 --- a/src/instructlab/sdg/pipeline.py +++ b/src/instructlab/sdg/pipeline.py @@ -34,7 +34,7 @@ def generate(self, dataset) -> Dataset: for block_prop in self.chained_blocks: block_type = block_prop["block_type"] block_config = block_prop["block_config"] - drop_columns = block_prop.get("drop_columns", None) + drop_columns = block_prop.get("drop_columns", []) gen_kwargs = block_prop.get("gen_kwargs", {}) drop_duplicates_cols = block_prop.get("drop_duplicates", False) block = block_type(**block_config) @@ -50,8 +50,9 @@ def generate(self, dataset) -> Dataset: dataset = block.generate(dataset, **gen_kwargs) + drop_columns_in_ds = [e for e in drop_columns if e in dataset.column_names] if drop_columns: - dataset = dataset.remove_columns(drop_columns) + dataset = dataset.remove_columns(drop_columns_in_ds) if drop_duplicates_cols: dataset = self._drop_duplicates(dataset, cols=drop_duplicates_cols) diff --git a/src/instructlab/sdg/utilblocks.py b/src/instructlab/sdg/utilblocks.py index 5f3c0407..e72c735b 100644 --- a/src/instructlab/sdg/utilblocks.py +++ b/src/instructlab/sdg/utilblocks.py @@ -10,12 +10,16 @@ class SamplePopulatorBlock(Block): - def __init__(self, config_paths, column_name, **batch_kwargs) -> None: - super().__init__(block_name=self.__class__.__name__) + def __init__(self, config_paths, column_name, post_fix="", **batch_kwargs) -> None: + super().__init__(block_name=self.__class__.__name__) # Call the base class's __init__ self.configs = {} for config in config_paths: + if post_fix: + config_name = config.replace(".yaml", f"_{post_fix}.yaml") + else: + config_name = config config_key = config.split("/")[-1].split(".")[0] - self.configs[config_key] = self._load_config(config) + self.configs[config_key] = self._load_config(config_name) self.column_name = column_name self.num_procs = batch_kwargs.get("num_procs", 8)