build-small-hackathon/job-search-distill

Job Search Distillation Corpus

A reasoning-trace SFT corpus for resume-aware job search. Teacher labels (search queries and fit evaluations, with full <think> 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 <think> 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 via JobSpy. The corpus is a sample, not an exhaustive snapshot of LinkedIn; rate-limit dynamics over-represent common, unrestricted query shapes.

Intended uses

• SFT distillation of small (≤10B) models on the query generation task (resume → LinkedIn search queries) or the fit evaluation task ((resume, job) → 5-dimension fit score with reasoning).
• Research on transferring reasoning traces from a teacher to a smaller student.
• Benchmark / evaluation set for resume-aware job retrieval and ranking systems.

Out-of-scope uses

• Production hiring decisions. The 5 fit-evaluation dimensions are author-defined and not a validated rubric. Treat total as an ordinal signal within a single candidate's shortlist, not as cross-candidate ground truth.
• Up-to-date job recommendations. jobs is a static snapshot; postings may have expired since scraping.
• PII-sensitive applications. Resumes are reproduced from the upstream source as-is, with no additional PII redaction.

Limitations & biases

• English only. Resume categories follow the US-centric livecareer.com taxonomy.
• Teacher errors propagate. Mis-categorizations, missed skills, and biased scoring by the teacher are present in the labels and will be learned by any student trained on them.
• Sample-biased job set. LinkedIn throttling shapes which queries successfully returned jobs; niche roles and restricted geographies are under-represented.

License

Apache-2.0, with these inheritances:

• Teacher labels (the reasoning and structured-output columns in query_gen_pairings and job_evals) are subject to DeepSeek's Open Platform Terms of Service.
• resume_corpus is redistributed from Divyaamith/Kaggle-Resume / livecareer.com — respect any upstream attribution requirements before commercial use.
• jobs is public LinkedIn data redistributed for research; downstream users are responsible for their own ToS and data-protection compliance.

Citation

If you use this dataset, please cite both this corpus and the upstream resume source.

@misc{job_search_distill,
  title = {Job Search Distillation Corpus},
  year  = {2026},
  url   = {https://huggingface.co/datasets/emrekuruu/job-search-distill}
}