|Redline — HR Interview Question Compliance Checker |
Category: HR , Model: Llama-3.3-70B-Instruct (full finetune) + Llama-4-Scout-17B-16E-Instruct (LoRA) — two architectures trained and compared
Problem
There is no public dataset for interview-question legal risk — companies that have asked risky interview questions have no incentive to publish them, and existing public guidance is unstructured listicle advice with no jurisdiction awareness. India and US employment law diverge in specific, checkable ways (e.g. asking prior salary is routine in India but banned in many US states/cities), and no existing tool reasons about that difference rather than pattern-matching on trigger words.
Solution
Given a job description, jurisdiction, and interview question, the model returns a risk classification (safe / legally_risky / bias_prone), the specific risk category, the legal basis (citing Title VII, ADEA, ADA, GINA, IRCA, Maternity Benefit Act 1961, POSH Act 2013, Rights of Persons with Disabilities Act 2016, Constitution Art. 15/16/17, NALSA v. UOI, Navtej Singh Johar v. UOI), a plain-language justification, and a safer reframe where applicable. Every label traces back to a hand-built legal reference table (30 categories, 15 per jurisdiction), not free-form LLM judgment. The seed set deliberately includes hard negatives — questions that sound risky but are legitimate when properly scoped (e.g. an ADA-compliant "can you lift 50 lbs, with or without accommodation" question).
Dataset
Primary seed (36 hand-verified rows) expanded to ~21,000 rows via AutoScientist's synthetic trajectory generation. A second, hand-written 40-row dataset was also used for training. A separate, genuinely held-out evaluation set (eval_set_v2, 30 rows, verified zero overlap with any training data) includes deliberate India/US jurisdiction flip-cases — the identical question posed in both countries, with different correct verdicts (e.g. salary-history questions: safe in India, legally risky in much of the US) — testing genuine cross-jurisdiction reasoning rather than keyword matching.
Results Model On-dataset win rate HR-category win rate Data quality grade Llama-3.3-70B, full FT 25 -> 75 18 -> 82 C -> B (+40.0%) Llama-4-Scout-17B, LoRA 28 -> 72 21 -> 79 D -> B (+117.5%) Limitations
Not legal advice — the reference table should be reviewed by qualified counsel before production use. Covers only India and US employment law.
Assets
Dataset cards (both training datasets + held-out eval), model card, live Gradio Space demo — all built. annotations_creators: [] language - - en language_creators: [] license- [] multilinguality: - monolingual pretty_name- 'adaption-interview_question_samples' size_categories: - 10K<n<100K source_datasets- - 'original' tags: - adaption - instruction-tuning - hr - career-workplace - other task_categories- [] task_ids: []
license: cc-by-4.0 language: - en - hi task_categories: - text-classification tags: - hr - compliance - legal - india - united-states - autoscientist-challenge-2026 pretty_name: HR Interview Question Compliance (Second Training Set) size_categories: - n<1K
HR Interview Question Compliance — Second Training Dataset
(interview_question_samples)
Dataset Summary
A 40-row hand-written dataset of interview questions labeled for legal/bias risk across
India and the US, used as a second training set for
adaption_llama_4_scout_17b_16_interview_question_sampl_1d90ba56
(Llama-4-Scout-17B-16E-Instruct, LoRA) via Adaption's AutoScientist. Hand-written rather
than synthetically expanded, this set adds high-precision, human-verified examples
alongside the larger synthetically-expanded primary training set
(hiring_question_compliance).
All generalization and benchmark claims for this project are reported against a separate,
independently held-out evaluation set (eval_set_v2.jsonl, 30 rows, zero overlap with any
training data) — see that dataset's card for details.
Schema
Same schema as the primary training dataset (hiring_question_compliance):
job_description, question, jurisdiction, risk_label, risk_category,
legal_basis, justification, reframe, plus an id field and a flip_case_group
field marking rows that form deliberate India/US jurisdiction flip-case pairs (identical
question, different correct verdict by jurisdiction).
Composition
- 40 rows, 19 US / 21 India (approx.)
- Includes 5 jurisdiction flip-case pairs
- Same legal reference table grounding as the primary training dataset
License
CC BY 4.0
Domain
- Hr (48%)
- Career-workplace (15%)
- Other (8%)
Language
- English (100%)
Tone
- Professional (63%)
- Informal (15%)
- Respectful (5%)
Evaluation Results
Quality Gains:
Grade Improvement:
Percentile Chart:
