spice-voice-lora

spice-voice-lora

A LoRA adapter that writes product descriptions in the Spices, Inc. house voice, given only a product name, category and ingredient list.

Trained as a deliberate test of a question rather than as a product: how much of a house style can a small model learn, and what does it fail to learn? The answer to the second half turned out to be the useful one.

What it does

Given bare facts:

Name: Satay Seasoning
Category: Seasoning Blends

it produces the full house format — an "also known as" line, an origin paragraph, a customer-type list, a Flavor Profile section and a How To Use section — at roughly the length and register of the published catalog.

What it gets wrong, and why that matters

It invents products and recipes that do not exist. Measured over 15 products, it fabricated 1.87 proper names per description: cross-sells to SKUs the company does not sell, and recipe titles appearing nowhere in the catalog.

This is not a training defect to be fixed with more steps. LoRA transfers behavior, not facts. The adapter learned that a How To Use section contains confident recipe names and produced confident recipe names.

Fabrication rose with fluency across all three arms tested:

The base model scores zero fabrications not because it is careful but because it writes nothing specific enough to be wrong — 2.87 contentless marketing phrases per description ("perfect for any occasion") and no names at all.

Do not use this to generate publishable copy unsupervised. Any real deployment needs retrieval over the actual SKU and recipe lists, with this adapter supplying only the voice.

Evaluation

Three arms on 15 products, judged by Claude Sonnet against a rubric measured from the 454-description corpus rather than written by hand — 96% carry a Flavor Profile section, 96% a How To Use section, 82% open with the product name, ~293 words average. A hand-written rubric would have scored the company's own copywriters as off-brand.

Caveat: the 15 evaluation products came from a superseded train/test split and appear in the training data, so this is not a clean held-out evaluation. Checked for memorization before reporting: output-to-published similarity is 0.075 mean, 0.157 max, against 0.026 for the base model — the adapter learned the template rather than the text. The scores are reported with that qualification.

Training

Data is the company's own published catalog copy, already public on the Spices, Inc. website. No customer data, no internal documents.

Usage

Prompts must use the Alpaca format it was trained on. Using a chat template instead produces incoherent output.

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "rsnyd/spice-voice-lora", max_seq_length=2048, load_in_4bit=True)
FastLanguageModel.for_inference(model)

alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.

### Instruction:
{}

### Input:
{}

### Response:
{}"""

inputs = tokenizer([alpaca_prompt.format(
    "Write a product description in the Spices, Inc. brand voice for the following blend.",
    "Name: Satay Seasoning\nCategory: Seasoning Blends",
    "",
)], return_tensors="pt").to("cuda")

print(tokenizer.batch_decode(model.generate(**inputs, max_new_tokens=600))[0])

Full writeup

Dataset construction, the notebook changes, the measured rubric, the fabrication detector and the fine-tune-vs-prompt-vs-RAG conclusions: github.com/rsnyd/cert-rag-cli/tree/main/finetune