distilabel: version: 1.5.3 pipeline: name: generate description: null steps: - step: name: load_data_from_hub_0 resources: replicas: 1 cpus: null gpus: null memory: null resources: null input_mappings: {} output_mappings: Job Role Description: anchor use_cache: false batch_size: 50 repo_id: dnth/ssf-dataset split: train config: null revision: null streaming: false num_examples: null storage_options: null runtime_parameters_info: - name: resources runtime_parameters_info: - name: replicas optional: true description: The number of replicas for the step. - name: cpus optional: true description: The number of CPUs assigned to each step replica. - name: gpus optional: true description: The number of GPUs assigned to each step replica. - name: memory optional: true description: The memory in bytes required for each step replica. - name: resources optional: true description: A dictionary containing names of custom resources and the number of those resources required for each step replica. - name: batch_size optional: true description: The number of rows that will contain the batches generated by the step. - name: repo_id optional: false description: The Hugging Face Hub repository ID of the dataset to load. - name: split optional: true description: The split of the dataset to load. Defaults to 'train'. - name: config optional: true description: The configuration of the dataset to load. This is optional and only needed if the dataset has multiple configurations. - name: revision optional: true description: The revision of the dataset to load. Defaults to the latest revision. - name: streaming optional: true description: Whether to load the dataset in streaming mode or not. Defaults to False. - name: num_examples optional: true description: The number of examples to load from the dataset. By default will load all examples. type_info: module: distilabel.steps.generators.huggingface name: LoadDataFromHub name: load_data_from_hub_0 - step: name: hard_triplets resources: replicas: 1 cpus: null gpus: null memory: null resources: null input_mappings: {} output_mappings: {} use_cache: true input_batch_size: 100 llm: generation_kwargs: temperature: 0.6 max_new_tokens: 512 use_offline_batch_generation: false offline_batch_generation_block_until_done: null jobs_ids: null model: gpt-4.1-mini base_url: https://api.openai.com/v1 default_headers: null max_retries: 6 timeout: 120 structured_output: null type_info: module: distilabel.models.llms.openai name: OpenAILLM group_generations: false add_raw_output: true add_raw_input: true num_generations: 1 use_default_structured_output: false triplet: true action: paraphrase hard_negative: true context: "\n\n## Task Overview\nYou are tasked with generating realistic job\ \ descriptions based on Singapore SkillsFuture Framework job descriptions.\ \ Your goal is to create both positive and a few negative examples for training\ \ a retrieval model.\n\n## Input Format\nYou will receive a job description\ \ from the Singapore SkillsFuture Framework containing:\n- Job title (e.g.,\ \ Audit Associate/Audit Assistant Associate)\n- Role responsibilities and\ \ duties\n- Work environment and supervision structure\n- Required skills\ \ and attributes\n- Professional conduct expectations\n\nThe text is a job\ \ description from the Singapore SkillsFuture Framework. Your task is to generate\ \ a realistic positive variation and a few hard negative variation.\n\n\n\ Hard negative is a job description that is similar in some ways but is ultimately\ \ incorrect for the given job. It must not be a simple paraphrase of a different,\ \ unrelated job.\n\nChoose from these strategies for hard negatives:\n1. **Different\ \ seniority level:** (e.g., Senior vs. Junior)\n2. **Different function:**\ \ (e.g., Business Valuation vs. Risk Management)\n3. **Different technology\ \ stack:** (e.g., Python/AWS vs. Scala/Azure for a Data Engineer role)\n4.\ \ **Different job title, same responsibilities:** (e.g., \"Web Developer\"\ \ vs. \"Frontend Engineer\")\n5. **Different industry jargon:** (e.g., \"\ Project Manager\" in Tech vs. \"Project Manager\" in Construction)\n6. **Broad\ \ vs. specialized role:** (e.g., a \"Data Scientist\" role for a \"Machine\ \ Learning Engineer\" positive)\n\nExample:\nJob Role: \"Data Analyst\"\n\ Positive: \"As a Data Analyst, you will interpret data and analyze results\ \ using statistical techniques. You will work to develop and implement data\ \ collection systems and other strategies that optimize statistical efficiency\ \ and quality.\"\n\n\nNegative: \"As a Senior Data Analyst, you will lead\ \ a team of junior analysts, manage end-to-end data projects, and present\ \ insights to key stakeholders to drive business strategy.\"\n\n \ \ \"As a Financial Analyst, you will analyze financial data to provide insights\ \ into business performance and support strategic decision-making.\"\n\n \ \ \"As a Healthcare Data Analyst, you will manage electronic health\ \ records and assist in clinical research data analysis.\"\n\n\nMake sure\ \ to mix a few strategies in one negative query. you could also use different\ \ strategies that is not mentioned above as long as it is a hard negative.\n\ Makesure to have a few negatives in one query like in the example given and\ \ make a new line for each new query.\nThe Job description of the negatives\ \ should give the real job description to a real job title.\nThe query should\ \ always include the job role and start with \"The \".\n" runtime_parameters_info: - name: resources runtime_parameters_info: - name: replicas optional: true description: The number of replicas for the step. - name: cpus optional: true description: The number of CPUs assigned to each step replica. - name: gpus optional: true description: The number of GPUs assigned to each step replica. - name: memory optional: true description: The memory in bytes required for each step replica. - name: resources optional: true description: A dictionary containing names of custom resources and the number of those resources required for each step replica. - name: input_batch_size optional: true description: The number of rows that will contain the batches processed by the step. - name: llm runtime_parameters_info: - name: generation_kwargs description: The kwargs to be propagated to either `generate` or `agenerate` methods within each `LLM`. keys: - name: max_new_tokens optional: true - name: logprobs optional: true - name: top_logprobs optional: true - name: echo optional: true - name: frequency_penalty optional: true - name: presence_penalty optional: true - name: temperature optional: true - name: top_p optional: true - name: stop optional: true - name: response_format optional: true - name: extra_body optional: true - name: use_offline_batch_generation optional: true description: Whether to use the `offline_batch_generate` method to generate the responses. - name: offline_batch_generation_block_until_done optional: true description: If provided, then polling will be done until the `ofline_batch_generate` method is able to retrieve the results. The value indicate the time to wait between each polling. - name: base_url optional: true description: The base URL to use for the OpenAI API requests. - name: api_key optional: true description: The API key to authenticate the requests to the OpenAI API. - name: default_headers optional: true description: The default headers to use for the OpenAI API requests. - name: max_retries optional: true description: The maximum number of times to retry the request to the API before failing. - name: timeout optional: true description: The maximum time in seconds to wait for a response from the API. - name: structured_output optional: true description: The structured output format to use across all the generations. - name: add_raw_output optional: true description: Whether to include the raw output of the LLM in the key `raw_output_` of the `distilabel_metadata` dictionary output column - name: add_raw_input optional: true description: Whether to include the raw input of the LLM in the key `raw_input_` of the `distilabel_metadata` dictionary column - name: num_generations optional: true description: The number of generations to be produced per input. type_info: module: distilabel.steps.tasks.sentence_transformers name: GenerateSentencePair name: hard_triplets - step: name: easy_triplets resources: replicas: 1 cpus: null gpus: null memory: null resources: null input_mappings: {} output_mappings: {} use_cache: true input_batch_size: 100 llm: generation_kwargs: temperature: 0.6 max_new_tokens: 512 use_offline_batch_generation: false offline_batch_generation_block_until_done: null jobs_ids: null model: gpt-4.1-mini base_url: https://api.openai.com/v1 default_headers: null max_retries: 6 timeout: 120 structured_output: null type_info: module: distilabel.models.llms.openai name: OpenAILLM group_generations: false add_raw_output: true add_raw_input: true num_generations: 1 use_default_structured_output: false triplet: true action: paraphrase hard_negative: false context: "\n\n\n## Task Overview\nYou are tasked with generating realistic job\ \ descriptions based on Singapore SkillsFuture Framework job descriptions.\ \ Your goal is to create both positive and a few negative examples for training\ \ a retrieval model.\n\n## Input Format\nYou will receive a job description\ \ from the Singapore SkillsFuture Framework containing:\n- Job title (e.g.,\ \ Audit Associate/Audit Assistant Associate)\n- Role responsibilities and\ \ duties\n- Work environment and supervision structure\n- Required skills\ \ and attributes\n- Professional conduct expectations\n\nThe text is a job\ \ description from the Singapore SkillsFuture Framework. Your task is to generate\ \ a realistic positive variation and a few easy negative variation.\n\nEasy\ \ negative is a job description and title that is completely irrelevant to\ \ the original job's industry, domain, or skills. It should be easy for the\ \ model to distinguish from the positive document.\n\nChoose from these strategies\ \ for easy negatives:\n1. Completely different industry (e.g., IT vs. Retail).\n\ 2. Unrelated skills and responsibilities.\n3. Completely different job function\ \ (e.g., Software Engineering vs. Human Resources).\n\nExample:\nJob Role:\ \ \"Data Analyst\"\n\nPositive: \"As a Data Analyst, you will interpret data\ \ and analyze results using statistical techniques. You will work to develop\ \ and implement data collection systems and other strategies that optimize\ \ statistical efficiency and quality.\"\n\nEasy Negative : \"As a Retail Manager,\ \ you will be responsible for overseeing daily store operations, managing\ \ inventory, and training staff to provide excellent customer service.\"\n\ \n \"As a Human Resources Assistant, you will provide administrative\ \ support to the HR department, assist with recruitment and onboarding processes,\ \ and maintain employee records.\"\n\n \"As a Professional\ \ Chef, you will be responsible for managing kitchen staff, creating new menu\ \ items, and ensuring all dishes are prepared to the highest quality standards.\"\ \n\n\nMake sure, that there is a variety of easy negatives across the dataset,\ \ using different strategies as outlined above.\nMakesure to have a few negatives\ \ in one query like in the example given.\nThe query should always include\ \ the job role and start with \"The \".\n" runtime_parameters_info: - name: resources runtime_parameters_info: - name: replicas optional: true description: The number of replicas for the step. - name: cpus optional: true description: The number of CPUs assigned to each step replica. - name: gpus optional: true description: The number of GPUs assigned to each step replica. - name: memory optional: true description: The memory in bytes required for each step replica. - name: resources optional: true description: A dictionary containing names of custom resources and the number of those resources required for each step replica. - name: input_batch_size optional: true description: The number of rows that will contain the batches processed by the step. - name: llm runtime_parameters_info: - name: generation_kwargs description: The kwargs to be propagated to either `generate` or `agenerate` methods within each `LLM`. keys: - name: max_new_tokens optional: true - name: logprobs optional: true - name: top_logprobs optional: true - name: echo optional: true - name: frequency_penalty optional: true - name: presence_penalty optional: true - name: temperature optional: true - name: top_p optional: true - name: stop optional: true - name: response_format optional: true - name: extra_body optional: true - name: use_offline_batch_generation optional: true description: Whether to use the `offline_batch_generate` method to generate the responses. - name: offline_batch_generation_block_until_done optional: true description: If provided, then polling will be done until the `ofline_batch_generate` method is able to retrieve the results. The value indicate the time to wait between each polling. - name: base_url optional: true description: The base URL to use for the OpenAI API requests. - name: api_key optional: true description: The API key to authenticate the requests to the OpenAI API. - name: default_headers optional: true description: The default headers to use for the OpenAI API requests. - name: max_retries optional: true description: The maximum number of times to retry the request to the API before failing. - name: timeout optional: true description: The maximum time in seconds to wait for a response from the API. - name: structured_output optional: true description: The structured output format to use across all the generations. - name: add_raw_output optional: true description: Whether to include the raw output of the LLM in the key `raw_output_` of the `distilabel_metadata` dictionary output column - name: add_raw_input optional: true description: Whether to include the raw input of the LLM in the key `raw_input_` of the `distilabel_metadata` dictionary column - name: num_generations optional: true description: The number of generations to be produced per input. type_info: module: distilabel.steps.tasks.sentence_transformers name: GenerateSentencePair name: easy_triplets - step: name: easy_keywords resources: replicas: 1 cpus: null gpus: null memory: null resources: null input_mappings: {} output_mappings: {} use_cache: true input_batch_size: 100 llm: generation_kwargs: temperature: 0.6 max_new_tokens: 512 use_offline_batch_generation: false offline_batch_generation_block_until_done: null jobs_ids: null model: gpt-4.1-mini base_url: https://api.openai.com/v1 default_headers: null max_retries: 6 timeout: 120 structured_output: null type_info: module: distilabel.models.llms.openai name: OpenAILLM group_generations: false add_raw_output: true add_raw_input: true num_generations: 1 use_default_structured_output: false triplet: true action: query hard_negative: false context: ' The text below is a job description from the Singapore SkillsFuture Framework. Your task is to generate two things: 1. A list of key skills for the job (the positive). 2. A list of unrelated skills (the easy negative). An **easy negative** is a set of keywords that is completely unrelated to the job description and its industry. It should be easy for the model to distinguish and serves to give the model a clear sense of what is definitively irrelevant. Example: Job Role: "Network Engineer" Positive Keywords: "Network administration, Cisco, firewall management, VPN" Easy Negative Keywords: "Retail inventory management, customer service, cash handling, visual merchandising" ' runtime_parameters_info: - name: resources runtime_parameters_info: - name: replicas optional: true description: The number of replicas for the step. - name: cpus optional: true description: The number of CPUs assigned to each step replica. - name: gpus optional: true description: The number of GPUs assigned to each step replica. - name: memory optional: true description: The memory in bytes required for each step replica. - name: resources optional: true description: A dictionary containing names of custom resources and the number of those resources required for each step replica. - name: input_batch_size optional: true description: The number of rows that will contain the batches processed by the step. - name: llm runtime_parameters_info: - name: generation_kwargs description: The kwargs to be propagated to either `generate` or `agenerate` methods within each `LLM`. keys: - name: max_new_tokens optional: true - name: logprobs optional: true - name: top_logprobs optional: true - name: echo optional: true - name: frequency_penalty optional: true - name: presence_penalty optional: true - name: temperature optional: true - name: top_p optional: true - name: stop optional: true - name: response_format optional: true - name: extra_body optional: true - name: use_offline_batch_generation optional: true description: Whether to use the `offline_batch_generate` method to generate the responses. - name: offline_batch_generation_block_until_done optional: true description: If provided, then polling will be done until the `ofline_batch_generate` method is able to retrieve the results. The value indicate the time to wait between each polling. - name: base_url optional: true description: The base URL to use for the OpenAI API requests. - name: api_key optional: true description: The API key to authenticate the requests to the OpenAI API. - name: default_headers optional: true description: The default headers to use for the OpenAI API requests. - name: max_retries optional: true description: The maximum number of times to retry the request to the API before failing. - name: timeout optional: true description: The maximum time in seconds to wait for a response from the API. - name: structured_output optional: true description: The structured output format to use across all the generations. - name: add_raw_output optional: true description: Whether to include the raw output of the LLM in the key `raw_output_` of the `distilabel_metadata` dictionary output column - name: add_raw_input optional: true description: Whether to include the raw input of the LLM in the key `raw_input_` of the `distilabel_metadata` dictionary column - name: num_generations optional: true description: The number of generations to be produced per input. type_info: module: distilabel.steps.tasks.sentence_transformers name: GenerateSentencePair name: easy_keywords - step: name: hard_keywords resources: replicas: 1 cpus: null gpus: null memory: null resources: null input_mappings: {} output_mappings: {} use_cache: true input_batch_size: 100 llm: generation_kwargs: temperature: 0.6 max_new_tokens: 512 use_offline_batch_generation: false offline_batch_generation_block_until_done: null jobs_ids: null model: gpt-4.1-mini base_url: https://api.openai.com/v1 default_headers: null max_retries: 6 timeout: 120 structured_output: null type_info: module: distilabel.models.llms.openai name: OpenAILLM group_generations: false add_raw_output: true add_raw_input: true num_generations: 1 use_default_structured_output: false triplet: true action: query hard_negative: true context: ' The text below is a job description from the Singapore SkillsFuture Framework. Your task is to generate two things: 1. A list of key skills for the job (the positive). 2. A list of unrelated skills (the easy negative). An **easy negative** is a set of keywords that is completely unrelated to the job description and its industry. It should be easy for the model to distinguish and serves to give the model a clear sense of what is definitively irrelevant. Example: Job Role: "Network Engineer" Positive Keywords: "Network administration, Cisco, firewall management, VPN" Easy Negative Keywords: "Retail inventory management, customer service, cash handling, visual merchandising" ' runtime_parameters_info: - name: resources runtime_parameters_info: - name: replicas optional: true description: The number of replicas for the step. - name: cpus optional: true description: The number of CPUs assigned to each step replica. - name: gpus optional: true description: The number of GPUs assigned to each step replica. - name: memory optional: true description: The memory in bytes required for each step replica. - name: resources optional: true description: A dictionary containing names of custom resources and the number of those resources required for each step replica. - name: input_batch_size optional: true description: The number of rows that will contain the batches processed by the step. - name: llm runtime_parameters_info: - name: generation_kwargs description: The kwargs to be propagated to either `generate` or `agenerate` methods within each `LLM`. keys: - name: max_new_tokens optional: true - name: logprobs optional: true - name: top_logprobs optional: true - name: echo optional: true - name: frequency_penalty optional: true - name: presence_penalty optional: true - name: temperature optional: true - name: top_p optional: true - name: stop optional: true - name: response_format optional: true - name: extra_body optional: true - name: use_offline_batch_generation optional: true description: Whether to use the `offline_batch_generate` method to generate the responses. - name: offline_batch_generation_block_until_done optional: true description: If provided, then polling will be done until the `ofline_batch_generate` method is able to retrieve the results. The value indicate the time to wait between each polling. - name: base_url optional: true description: The base URL to use for the OpenAI API requests. - name: api_key optional: true description: The API key to authenticate the requests to the OpenAI API. - name: default_headers optional: true description: The default headers to use for the OpenAI API requests. - name: max_retries optional: true description: The maximum number of times to retry the request to the API before failing. - name: timeout optional: true description: The maximum time in seconds to wait for a response from the API. - name: structured_output optional: true description: The structured output format to use across all the generations. - name: add_raw_output optional: true description: Whether to include the raw output of the LLM in the key `raw_output_` of the `distilabel_metadata` dictionary output column - name: add_raw_input optional: true description: Whether to include the raw input of the LLM in the key `raw_input_` of the `distilabel_metadata` dictionary column - name: num_generations optional: true description: The number of generations to be produced per input. type_info: module: distilabel.steps.tasks.sentence_transformers name: GenerateSentencePair name: hard_keywords connections: - from: load_data_from_hub_0 to: - hard_triplets - easy_triplets - easy_keywords - hard_keywords - from: hard_triplets to: [] - from: easy_triplets to: [] - from: easy_keywords to: [] - from: hard_keywords to: [] routing_batch_functions: [] type_info: module: distilabel.pipeline.local name: Pipeline requirements: []