drug-screening-llm

DrugScreenLLM - A Domain-Specific LLM for Workplace Drug Screening Compliance

A QLoRA fine-tune of Meta's Llama 3.2 3B Instruct, specialized for workplace drug screening compliance. Built for HR professionals, occupational health nurses, Medical Review Officers (MROs), and employers navigating SAMHSA, DOT, FDA, and state-level workplace drug testing regulation.

This repository hosts the V2 adapter (citation- and refusal-augmented). All numbers below are V2 and match the accompanying paper; see Version history for the V1/V2 distinction.

Problem Statement

Workplace drug screening operates under a complex, frequently updated regulatory framework spanning SAMHSA mandatory guidelines, DOT 49 CFR Part 40, FDA device regulations, and 50 varying state marijuana laws. Misinterpretation of cutoff thresholds, panel compositions, or jurisdiction-specific rules can result in legal liability, wrongful termination claims, or compliance violations.

Intended Use

Intended users: HR professionals managing workplace drug-free policies; occupational health nurses interpreting screening results; Medical Review Officers; employers in DOT-regulated industries; compliance officers.

Intended use cases: answering regulatory compliance questions (SAMHSA, DOT, FDA); cutoff thresholds by drug class and specimen type; state-specific marijuana testing laws; the two-step screening and confirmation process; MRO and split-specimen procedures.

Out of scope:

• NOT for clinical diagnosis
• NOT a substitute for legal advice
• NOT for individual employee medical decisions
• Results should always be verified against current Federal Register notices

Training Data

Fine-tuned on the Drug Screening Compliance Benchmark Dataset - 710 expert-crafted Q&A pairs grounded exclusively in publicly available regulatory and peer-reviewed sources:

• SAMHSA Mandatory Guidelines for Federal Workplace Drug Testing Programs (Urine, 2017; Oral Fluid, 2023)
• DOT 49 CFR Part 40 (current)
• FDA drug testing device regulations
• StatPearls peer-reviewed toxicology literature (NCBI Bookshelf)
• Federal Register notices (2024-2025), including the fentanyl panel addition effective July 7, 2025
• State marijuana workplace testing laws (all 50 states)

Splits: 639 training / 71 validation. The validation split contains 67 on-topic regulatory questions and 4 off-topic refusal examples. Of the 710 total, 40 are off-topic refusal pairs added during V2 augmentation.

All training data is traceable to primary public sources. No proprietary or confidential data was used. The dataset is released under CC BY-NC 4.0.

Training Approach

Training metrics (V2)

Evaluation Results

Evaluated with an LLM-as-judge protocol using openai/gpt-oss-120b (served via Groq) as an independent judge - a different model family from every candidate. Each dimension is scored 1-5. Correctness, citations, and on-topic refusal are scored over the 67 on-topic validation questions; off-topic refusal over 10 held-out off-topic prompts. Candidate generation used temperature 0.3 / top-p 0.9; the judge ran at temperature 0, one judging pass per item.

The overall score approaches that of a 70B reference model at roughly 1/23 the parameter count, driven substantially by off-topic refusal calibration (2.4 -> 4.7 from V1 to V2). Correctness remains below the frontier model and declined slightly from V1 (2.940 -> 2.761) - a specialization trade-off reported in full in the paper rather than omitted here.

Version history

• V1 - baseline fine-tune, 670 examples, no explicit citation formatting. Overall 2.932, correctness 2.940, citations 1.388, off-topic refusal 2.4.
• V2 (current) - adds explicit source citations to all answers plus 40 off-topic refusal examples (710 total). Trained fresh from the base model, not continued from the V1 adapter.

Limitations

• Evaluation set: the 71-example validation split was also used for checkpoint selection, so reported scores are not from a fully held-out test set.
• Evaluation scale: off-topic refusal is measured on only 10 prompts; those scores are indicative rather than statistically precise. Scoring used a single judging pass.
• Judge validation: judge-human agreement was not separately validated for this judge model or this regulatory domain.
• Correctness: below the 70B reference model; refusal calibration alone does not establish production readiness.
• Regulatory cutoffs and panel compositions change frequently - always verify against the current Federal Register notice.
• State marijuana laws evolve rapidly - consult current state statutes before making employment decisions.
• Reflects training data as of June 2026 and will not capture later regulatory changes.

Regulatory Notice

Drug screening decisions with legal or employment consequences must be reviewed by a qualified Medical Review Officer (MRO) and legal counsel. This model is an informational tool only.

License

This adapter is derived from Meta's Llama 3.2 3B Instruct and is therefore governed by the Llama 3.2 Community License, not an OSI-approved open source license. The accompanying dataset is released under CC BY-NC 4.0 (non-commercial). Review both before use.

Citation

@misc{drugscreenllm2026,
  title  = {DrugScreenLLM: A Domain-Specific Language Model for Workplace Drug Screening Compliance},
  author = {Bhattacharjya, Sairath},
  year   = {2026},
  publisher = {HuggingFace},
  url    = {https://huggingface.co/oikyoni/drug-screening-llm}
}

Author

Developed by Sairath Bhattacharjya, who brings professional experience in federally regulated workplace drug testing compliance.