aznatkoiny/cannabis-fda-extractive-pilot

FDA Cannabis Extractive Experimental Pilot

Experimental, automatically screened, unreviewed draft dataset. This dataset is not medical advice, is not production-ready, and must not be represented as clinician-reviewed, legally cleared, or suitable for patient-facing systems.

This small English conversational dataset was created to test an auditable Gemma 4 fine-tuning pipeline. It contains exact answer passages from captured FDA pages about CBD/cannabis safety, paired with source-grounded questions. It does not contain strain reviews, patient experiences, customer reviews, dispensary catalogs, individualized treatment advice, or evidence about what a particular person should expect from a strain.

No Leafly, Weedmaps, AllBud, Reddit, Kaggle review mirror, or scraped consumer-review site content was used.

Composition

The 95 tuning/validation rows represent only 19 unique assistant answers. Prompt wrappers add surface-form variety, not new facts.

Files

• —train.jsonl: 80 conversational draft records.
• —validation.jsonl: 15 conversational draft records.
• —locked_eval_cases.jsonl: five page-held-out draft eval cases with lexical concept rubrics.
• —split_groups.json: deterministic source-group split receipt.
• —pilot_manifest.json: self-hashed transformation, source, count, limitation, and automated-check receipt.
• —fda-extractive-candidates-v1.json: exact extraction recipe and source locators.
• —fda-extractive-v1.sources.csv: candidate source registry; training permission remains false under the production gate.
• —fda-extractive-v1.build.json: upstream candidate-build receipt.
• —acquisition-receipt.json: hashes for captured FDA source evidence.
• —SHA256SUMS: hashes for this hosted package, excluding itself.

Loading

The tuning and evaluation records have intentionally different schemas, so they are separate configurations:

from datasets import load_dataset

tuning = load_dataset(
    "aznatkoiny/cannabis-fda-extractive-pilot",
    "tuning",
)
evaluation = load_dataset(
    "aznatkoiny/cannabis-fda-extractive-pilot",
    "evaluation",
)

Tuning rows use a strict chat-record schema with messages, source IDs, risk tags, hashes, and draft-review fields. Evaluation rows use prompts plus required/forbidden concept patterns and separate source IDs.

Source and transformation approach

The builder accepts only allowlisted FDA pages and a captured FDA website-policy page. It verifies exact bytes against acquisition-receipt.json, extracts normalized visible article text, and requires every assistant answer to occur exactly once in its specified page. Images, logos, scripts, linked third-party content, stale numeric incident counts, phone numbers, animal-only material, dosing advice, strain claims, and customer reviews were excluded.

Assistant answers were preserved verbatim. Four project-authored wrapper prefixes were added around each retained base prompt. Each five-variant group remains entirely in train or validation.

Automated checks included:

• —raw source, recipe, receipt, and output hashes;
• —exact answer/source matching;
• —unique record IDs and content hashes;
• —train/validation group isolation;
• —train/eval page and source-ID isolation;
• —prompt-similarity screening;
• —exact, substring, contiguous-token, and required-concept leakage checks;
• —phone/email pattern screening; and
• —byte-for-byte deterministic reconstruction.

Three training anchors were excluded before expansion because they could teach held-out answer concepts. The maximum remaining contiguous train/eval answer overlap was five normalized tokens; the rejection threshold was eight.

Review and release status

The canonical manifest deliberately says:

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

Those values are preserved. publish_allowed=false means the project's automated release pipeline did not approve these records for publication or deployment. Public hosting here is a transparent research copy made at the project owner's explicit request; it does not become human-reviewed or production-approved because it is downloadable.

The records declare CC0-1.0 and are derived from U.S. federal government FDA pages under the captured website policy, but no independent legal professional completed the project's formal rights-clearance workflow. Users must perform their own review before redistribution or model training.

Known limitations

• —Nineteen unique answers are too few for a useful domain assistant.
• —The corpus overrepresents federal FDA safety language and does not capture the broader evidence base.
• —It contains no strain-level chemistry, batch-level COAs, route/dose context, or consented experience reports.
• —It cannot support claims about expected subjective effects, efficacy, personalized risk, or product quality.
• —Source snapshots and regulatory statements can become stale.
• —Prompt wrappers can make naive row counts look more diverse than the corpus is.
• —The five-case eval set is tiny, lexical, and not clinician-authored.
• —Automated screening does not replace medical, scientific, privacy, legal, or data-rights review.
• —Training on this data can still produce hallucinations, omissions, or unsafe answers.

Observed model result

The companion adapter `aznatkoiny/gemma-4-e4b-cannabis-fda-pilot` trained for one epoch and scored 0.50 against a required 0.80 on the five held-out cases. It matched zero complete required-concept sets. This dataset therefore demonstrates pipeline mechanics only, not model readiness.

Improvement roadmap

1. 1.Obtain explicitly licensed or consented first-party strain-experience contributions with privacy review, contributor terms, moderation, and withdrawal controls.
2. 2.Add permissively licensed structured cannabinoid/terpene chemistry and batch COA data. Keep changing product facts in citation-bearing retrieval rather than model weights.
3. 3.Record dose, route, timing, tolerance, concurrent substances, and adverse outcomes using controlled schemas without turning correlations into recommendations.
4. 4.Deduplicate and split by contributor, product, chemotype, batch, and source to prevent participant or near-duplicate leakage.
5. 5.Commission independent medical, scientific, privacy, legal, and data-rights review.
6. 6.Expand to at least 200 human-reviewed safety/evidence cases, including drug interactions, pregnancy, minors, impairment, overconsumption, dosing refusal, uncertainty, citations, and adversarial prompts.
7. 7.Compare the pinned base and adapter, retain complete inference manifests, run blinded human evaluation, and require no safety-category regression.
8. 8.Fine-tune again only after unique factual and behavioral diversity materially increases.

License and non-endorsement

Dataset records and project-authored prompt wrappers are offered under CC0-1.0. Source provenance and captured policy evidence are documented in the receipts. No FDA endorsement is claimed. References to FDA identify the source of public information, not sponsorship or approval of this dataset or any model trained from it.