This model is tuned to answer questions about cannabis strain profiles, reported effects, flavors/aromas, terpene signals, cannabinoid signals, lineage, and grow notes.

It should not be used as a source of medical, legal, or safety-critical advice. The training data is based on crowdsourced strain records, so outputs should be treated as consumer-reported tendencies rather than clinical evidence.

Training Data

Generated instruction data from Hormold/leafly-full-dump-cannabis:

24,136 training examples
400 evaluation examples
Fields used: strain names, categories, descriptions, reported effects, flavors, terpenes, cannabinoids, ratings, lineage, grow fields, symptom/condition aggregates, and negative-effect signals.

Recommended System Prompt

You are a rigorous cannabis strain specialist. Give professional, evidence-aware, consumer-safe answers about strain taxonomy, reported effects, aromas, terpenes, cannabinoids, lineage, and practical selection. Do not present crowdsourced strain reports as medical proof, do not give medical or legal advice, and mention uncertainty when data is sparse.

Prompt Format

This model was trained with this lightweight chat format:

text

<system>
You are a rigorous cannabis strain specialist. Give professional, evidence-aware, consumer-safe answers about strain taxonomy, reported effects, aromas, terpenes, cannabinoids, lineage, and practical selection. Do not present crowdsourced strain reports as medical proof, do not give medical or legal advice, and mention uncertainty when data is sparse.
</system>
<user>
Give me an expert strain profile for Blue Dream.
</user>
<assistant>

Example

python

from transformers import AutoTokenizer, AutoModelForCausalLM

repo_id = "ZeroSaturn/gemma4-e2b-cannabis-strain-expert"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id, device_map="auto", torch_dtype="auto")

messages = [
    {"role": "system", "content": 'You are a rigorous cannabis strain specialist. Give professional, evidence-aware, consumer-safe answers about strain taxonomy, reported effects, aromas, terpenes, cannabinoids, lineage, and practical selection. Do not present crowdsourced strain reports as medical proof, do not give medical or legal advice, and mention uncertainty when data is sparse.'},
    {"role": "user", "content": "Compare Blue Dream and Granddaddy Purple for reported effects and aroma."},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=500, temperature=0.7, top_p=0.9)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Training Setup

Base model: principled-intelligence/gemma-4-E2B-it-text-only
Adapter source: outputs/gemma4-e2b-cannabis-qlora
Method: QLoRA / LoRA
LoRA rank: 16
Trainable params: about 24.2M
Max sequence length: 2048
Steps: 700

Safety Notes

Cannabis laws vary by jurisdiction. Effects vary by person, dose, route, tolerance, and batch chemistry. This model should qualify strain claims and recommend checking current lab results when precision matters.