emrekuruu/job-search-distill · Datasets

emrekuruu/job-search-distill · Datasets at Hugging Face Job Search Distillation Corpus A reasoning-trace SFT corpus for resume-aware job search. Teacher labels (search queries and fit evaluations, with full` ` reasoning preserved) generated by DeepSeek V4 Pro. Four relational configs cover the full pipeline: resumes → search queries → scraped jobs → fit evaluations. Dataset structure | Config | Contents | | --- | --- | | `resume_corpus` | `resume_id`,`category`,`resume` | | `query_gen_pairings` | `resume_id`, teacher`reasoning`, list of`queries`(each with`query_id`) | | `jobs` | `job_id`,`resume_id`,`query_id`,`title`,`company`,`location`,`description`,`job_url` | | `job_evals` | `eval_id`,`resume_id`,`job_id`, teacher`reasoning`, 5`dimensions`(each 0–20),`total`(0–100),`overall_reasoning` | Keys:`resume_id` is the integer PK.`query_id = "{resume_id}-{idx}"` indexes into the resume's`queries` list.`job_id = sha256(job_url):16`. ``` from datasets import load_dataset queries = load_dataset("emrekuruu/job-search-distill", "query_gen_pairings", split="train") evals = load_dataset("emrekuruu/job-search-distill", "job_evals", split="train") ``` Distillation process - Teacher model:`deepseek-v4-pro` via the DeepSeek Platform API. - Sampling: temperature`1.0`, max tokens`16,000`, request timeout`600 s`. - Reasoning capture: the teacher's` ` content is stored verbatim in the`reasoning` columns, alongside the structured label. - Output validation: every label is Pydantic-validated against a fixed schema before being written; rows that fail validation are dropped. - Random seed:`42`, used for query sampling and any downstream splitting. The exact prompts and generation code are at github.com/build-small-hackathon/job-search. Source data - Resumes —`resume_corpus` is built from Divyaamith/Kaggle-Resume, originally sourced from livecareer.com. Categories follow the upstream taxonomy. - Jobs —`jobs` are public LinkedIn job postings retrieved v