README.md

---
license: mit
language: [en]
task_categories: [text-classification]
tags: [resume, cv, skills, hr-tech, nli, synthetic]
pretty_name: SkillJev data
size_categories: [10K<n<100K]
configs:
- config_name: default
  data_files:
  - {split: train, path: train.jsonl}
  - {split: heldout_locked, path: heldout_locked.jsonl}
---
# SkillJev data

This dataset holds (job requirement, candidate evidence) pairs with three labels: **skill group**, **evidence** (`supported/partial/contradicted/insufficient`), and **proficiency** (`0–5` or `unknown`; unknown ≠ 0). It was used to train [`niko0xdev/skilljev`](https://huggingface.co/niko0xdev/skilljev).

## Files
| File | Rows | Description |
|---|---|---|
| `train.jsonl` | 38,938 | training set (v008, after PII filtering) |
| `heldout_locked.jsonl` | 90 | locked evaluation set, byte-identical to the version used for every reported metric (`heldout_locked_v003.sha256`) |
| `trainer_train.jsonl`, `trainer_heldout_locked.jsonl` | 116,814 / 270 | same data in multiple-choice trainer format `{state, question, options, type, target, meta}`, 3 rows per sample (one per head) |
| `RUBRIC.md`, `SCHEMA.md`, `TAXONOMY.md` | – | labeling rubric (incl. Real-CV convention v0.2), JSON schema, skill taxonomy (55 skills, 10 groups) |

## Composition (train)
| Source (`meta.source`) | Rows | Labels (`meta.label_status`) |
|---|---|---|
| synthetic (`synthetic_v002/v004/v005/v006/v008`) | 31,960 | template-generated: case types × rubric anchors × paraphrases, including level 3/4/5 boundary, partial-vs-supported contrast, and convention-v0.2 cases |
| `djinni_cv_en` real CV excerpts | 6,928 | 1,223 `agent_reviewed_v005` (read individually and relabelled per rubric); 5,705 `silver_rules_v006` (rule-based labels, about 76% agreement with reviewed labels); convention v0.2 applied to all |
| `handwritten_v001` | 50 | hand-written draft labels |

**Held-out:** 70 synthetic samples use templates that never appear in train; 20 are reviewed real excerpts. The CV records they come from are excluded from train, and train is deduplicated against held-out text.

## Sources & attribution
Real excerpts are single sentences taken from [**lang-uk/recruitment-dataset-candidate-profiles-english**](https://huggingface.co/datasets/lang-uk/recruitment-dataset-candidate-profiles-english) (Djinni candidate profiles, **MIT license**). Please cite that dataset as well when using these excerpts. Synthetic text was written from templates for this project.

## Anonymization
- Only short excerpts are kept (≤ ~260 characters, one sentence or line). No full profiles and no source profile IDs beyond the dataset's own record id (`meta.source_record`).
- Excerpts were filtered at extraction and rescanned before publishing. Any row matching an email, phone number, URL or domain, social handle (@, LinkedIn, GitHub, Telegram, Skype), self-introduction ("my name is"), or a list of common first names was **dropped** (103 rows removed from train, 0 from held-out).
- Residual risk: company or project names can remain, and name detection is heuristic. Report issues via the Community tab and the rows will be removed.

## Labels
See `RUBRIC.md`. Labels are synthetic or agent/rule-produced and have **not** been verified by a professional recruiter. Use this data for research and prototyping, not as ground truth for hiring decisions.