aznatkoiny/gemma-4-e4b-cannabis-fda-pilot

Gemma 4 E4B FDA Cannabis Safety Pilot

Experimental research artifact — not medical advice and not production-ready.

Do not use this adapter to recommend cannabis products, predict a person's response, diagnose or treat a condition, determine dosing, or replace a clinician, pharmacist, Poison Control, or emergency services. It has not been reviewed by medical experts and failed its held-out evaluation.

This is a PEFT/QLoRA adapter for google/gemma-4-E4B-it. It was created to test an auditable fine-tuning pipeline using a very small FDA-derived safety dataset. It is not a strain-experience model: it contains no customer reviews, dispensary catalog, strain chemistry, clinical outcome, or personalized patient data.

The exact base-model revision used for training and evaluation was:

google/gemma-4-E4B-it@ee0ef6023621cff504d758262d4e04895a5af4a2

Release status

The attached receipts deliberately retain:

experimental=true
automated_checks_only=true
human_review_performed=false
rights_clearance_performed=false
publish_allowed=false
production_release_allowed=false
production_use_allowed=false

publish_allowed=false is the result of this project's automated production-release gate. It means the pipeline did not approve this artifact for publication or deployment; placing a copy on a model host for research transparency does not turn that failed gate into approval. The artifact must not be represented as medically reviewed, release-ready, or suitable for patient-facing use.

Training data

The pilot dataset, fda-extractive-pilot-v1, was deterministically derived from captured FDA web pages about CBD products and foods, delta-8 THC, accidental ingestion by children, pregnancy/breastfeeding, and federal cannabis-product regulation. It has:

• 19 retained, unique FDA extractive answer groups from five training pages;
• five project-authored prompt wrappers per answer, with answers left unchanged;
• 80 training rows and 15 grouped validation rows;
• five untouched evaluation cases from a sixth, page-disjoint FDA source; and
• three source drafts excluded before tuning.

The row count therefore overstates semantic diversity: 95 tuning rows represent only 19 unique answers. Automated checks covered hashes, group separation, overlap, prompt similarity, leakage anchors, and phone/email patterns. They are not substitutes for human content, legal, scientific, or medical review. See dataset_manifest.json.

No Leafly, Weedmaps, AllBud, Reddit, scraped customer reviews, or repackaged mirrors were used. Do not add such material without explicit rights covering collection, model training, derived weights, redistribution, and the intended use.

Training

Training ran for one epoch on an NVIDIA GB10 using 4-bit NF4 QLoRA:

Setting	Value
Base revision	ee0ef6023621cff504d758262d4e04895a5af4a2
Epochs	1
Maximum sequence length	512
Train/eval batch size	4 / 4
Gradient accumulation	1
Learning rate	2e-4
LoRA	r=8, alpha=16, dropout=0.05, all linear modules
Quantization	NF4, double quantization
Gradient checkpointing	enabled
Seed	42
Training loss	1.9135 (1.9134718984365464 unrounded)
Runtime	37.4408 seconds, 20 optimizer steps

The tested package versions are recorded in training_run.json: Transformers 5.14.1, PEFT 0.19.1, PyTorch 2.13.0, TRL 1.9.0, bitsandbytes 0.49.2, Accelerate 1.14.0, and Datasets 5.0.0.

Evaluation: failed

The adapter scored 0.50 on five held-out, page-disjoint FDA cases and the report records passed=false. All five generations were non-empty and there were no runtime errors or rubric-critical failures, but every response failed to cover the required source concepts. The score should not be read as 50% medical accuracy: each response received credit for merely being present, while its required-concept score was zero.

Both tested categories failed:

Category	Cases	Mean score	Passed
Medical claims	3	0.50	No
Minors	2	0.50	No

See evaluation_report.json for prompts, responses, criteria, and exact hashes.

Loading the adapter

The example pins the base model to the exact training revision. Replace ADAPTER_ID with this repository's Hub ID (or a local clone). Gemma access may require accepting Google's terms on the hosting service.

import torch
from peft import PeftModel
from transformers import AutoModelForMultimodalLM, AutoProcessor

ADAPTER_ID = "aznatkoiny/gemma-4-e4b-cannabis-fda-pilot"
BASE_MODEL = "google/gemma-4-E4B-it"
BASE_REVISION = "ee0ef6023621cff504d758262d4e04895a5af4a2"

# The adapter repository includes the processor/tokenizer/chat template used in training.
processor = AutoProcessor.from_pretrained(ADAPTER_ID)
base_model = AutoModelForMultimodalLM.from_pretrained(
    BASE_MODEL,
    revision=BASE_REVISION,
    dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base_model, ADAPTER_ID)
model.eval()

messages = [{"role": "user", "content": "What are important limitations of CBD safety data?"}]
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=False,
).to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
        num_beams=1,
    )

generated = output[0][inputs["input_ids"].shape[-1] :]
decoded = processor.decode(generated, skip_special_tokens=False)
print(processor.parse_response(decoded, prefix=inputs["input_ids"]))

Even for research, independently verify every factual statement against current, authoritative sources. The example is an inference demonstration, not an endorsement of the generated answer.

Included artifacts and primary hashes

File	Purpose	SHA-256
adapter_model.safetensors	LoRA adapter weights	f46e9446ff25989d8a6300f338c51c56823734fefd792193231bd2437e3c975f
adapter_config.json	PEFT configuration	89273cdebde38a1bd4cdb5ea5f9b0f89b6736aeedaacd1749c748790be8a3958
chat_template.jinja	Exact chat template	0a2c8073c878ab1da004bee933a998606537bbb62016310352c7285c3f01c5b5
tokenizer.json	Tokenizer	0d1d10b5165f3543ad0acbe0ad21105552c7a179fbb90a0d4b73ea22513f954a
training_run.json	Training receipt and gate state	cc3f6fae5011dc7852b4ad399fd4b13cfbaed0f4481ea7454fadf97819083a91
dataset_manifest.json	Dataset construction receipt	f4e36a009f1aba796bdfe2e768f907932b914fb8d2207bc0d4365344734489e0
evaluation_report.json	Failed held-out evaluation	1e3bceae5338d13002e96b71e126d3ee104a11095dc9b4347f74cb482e152acf

SHA256SUMS covers the complete staged package. Base-model weights and training_args.bin are not included.

Limitations

• This adapter learned a narrow FDA safety style, not strain effects or patient experiences.
• Nineteen unique answers are far too few for broad capability or reliable safety.
• Prompt wrappers increase formatting variety, not evidence diversity.
• Source extracts and evaluation criteria received automated checks only.
• The corpus is a dated snapshot; regulations, evidence, and FDA guidance can change.
• The evaluation is tiny, lexical, and not clinician-authored. It does not establish factuality, calibration, fairness, robustness, or medical safety.
• Fine-tuning cannot guarantee citations, prevent hallucinations, or make an LLM appropriate for high-stakes decisions.
• Cannabis effects vary with composition, dose, route, tolerance, interactions, individual physiology, and many other factors not represented here.

Future work

1. Complete independent human medical, scientific, data-rights, and legal review.
2. Add substantially more unique, source-grounded examples; do not manufacture scale by repeating prompt wrappers.
3. Obtain explicitly licensed or consented first-party experience data and structured, quality-controlled chemistry/COA data. Track provenance at record level and exclude identifying information.
4. Split by source, product, batch, and contributor to prevent near-duplicate or participant leakage.
5. Keep changing facts in a citation-bearing retrieval layer rather than model weights; add freshness checks and an authoritative-source fallback.
6. Build clinician-reviewed evaluations for contraindications, drug interactions, pregnancy, minors, impairment, emergencies, uncertainty, abstention, unsupported medical claims, and adversarial prompts.
7. Compare the adapter against the pinned base model, use blinded human evaluation, test calibration and subgroup behavior, and publish confidence intervals.
8. Repeat tuning across seeds and hyperparameters only after data quality improves, then require every production release gate to pass before deployment.

License and attribution

The adapter package is offered under Apache-2.0; see LICENSE and NOTICE. The Gemma 4 base model is not included and remains subject to its own Apache-2.0 materials and notices. FDA-derived records are documented separately as CC0/public domain candidates, subject to the qualifications in THIRD_PARTY_NOTICES.md and the fact that this pilot did not complete human rights clearance. No FDA endorsement is claimed.