Abercrombie-Decision

An Unsloth decision model that classifies trademarks on the Abercrombie distinctiveness spectrum (Generic / Descriptive / Suggestive / Arbitrary / Fanciful). It is a LoRA adapter on unsloth/Qwen3.5-4B plus a Clef-style decision head.

It does not generate text. The LLM reads the mark once, and the head scores the options for six questions together: five yes/no doctrine sub-questions and a final 5-level distinctiveness score. Every answer comes with a probability.

This is a proof of concept and the companion to DoodDood/abercrombie-grpo. It uses the same base model, training marks and test set, with a different training method.

Results

On the 95-row LegalBench Abercrombie test set. Training data excluded every test mark.

Category	Base Qwen3.5-4B	Abercrombie-GRPO	This model (seed 42)	Mean of 3 seeds
Generic	89%	100%	100%	100%
Descriptive	100%	74%	95%	95%
Suggestive	5%	26%	26%	25%
Arbitrary	0%	47%	68%	65%
Fanciful	5%	95%	89%	91%
Overall	40.0%	68.4%	75.8%	75.1% +/- 1.2

Mean ordinal distance: 0.34 (GRPO 0.53, base 1.09). Training took about 7 minutes on one RTX PRO 6000 (Colab G4).

Sub-question accuracy (seed 42): coined 98%, semantic link 93%, imagination 82%, conveys 81%, genus 95%. The sub-answers, run through the routing rule below, agree with the model's own final answer on 94% of marks.

A single-question variant (distinctiveness only) scored 74.4% +/- 3.7, which is not a significant difference (paired sign test p = 0.50).

Limitations (read before using)

• Small test set. 95 rows means each run's accuracy carries roughly +/- 9 points of uncertainty.
• Suggestive marks are weak (about 25%). They mostly get called Descriptive (Coppertone, Chicken of the Sea, Orange Crush).
• Famous brands get pulled toward Generic. Microsoft, Netflix, Airbus and Amazon are misclassified in every run. All training marks are synthetic and invented, so the model has never seen a famous real brand.
• Overconfident on real marks. Probabilities were calibrated on a synthetic holdout (ECE 0.008) but are overconfident on LegalBench (ECE 0.145 for this seed, about 0.16 averaged over seeds). Treat the probabilities as a ranking, not as calibrated odds.
• Sub-question labels are derived, not annotated. The training data labels only the final category. Sub-question gold was derived from the routing rule plus fixed conventions (for example, a generic name "conveys" its goods).
• Not legal advice. This is a research demo, not a clearance or registrability opinion.

Usage

Needs a GPU and Unsloth (trained with unsloth==2026.10.3). It loads with FastDecisionModel, not a transformers text-generation pipeline.

pip install unsloth==2026.10.3 unsloth_zoo==2026.10.3

from unsloth import FastDecisionModel

model, tokenizer = FastDecisionModel.from_pretrained("DoodDood/abercrombie-decision")
FastDecisionModel.for_inference(model)

CATEGORIES = ["Generic", "Descriptive", "Suggestive", "Arbitrary", "Fanciful"]
QUESTIONS = {
    "coined": {"type": "noul", "instructions": "Is the mark an invented term created solely for trademark use, with no prior independent meaning?"},
    "semantic_link": {"type": "noul", "instructions": "Does the mark's ordinary dictionary meaning have any plausible semantic relationship to the goods or services?"},
    "imagination": {"type": "noul", "instructions": "Must the consumer use imagination, thought, or a multi-step mental process to connect the mark to the nature of the goods or services?"},
    "conveys": {"type": "noul", "instructions": "Does the mark immediately convey an idea of a feature, quality, function, ingredient, or characteristic of the goods or services to the relevant purchasing public?"},
    "genus": {"type": "noul", "instructions": "Does the relevant purchasing public understand the mark primarily as the name of the general category of goods or services, rather than as an indicator of source?"},
    "distinctiveness": {
        "type": "score",
        "instructions": "Where does the mark fall on the Abercrombie spectrum of trademark distinctiveness for these goods or services?",
        "criteria": [
            "Generic: the common name of the goods or services; never protectable.",
            "Descriptive: directly describes a feature, quality or purpose of the goods.",
            "Suggestive: hints at the goods but needs imagination to connect.",
            "Arbitrary: a real word with no relationship to the goods.",
            "Fanciful: an invented word with no prior meaning.",
        ],
    },
}

# Input format: `The mark "X" for Y.` (the LegalBench phrasing it was trained on)
answers = FastDecisionModel.predict(model, tokenizer, 'The mark "Kodak" for cameras.', QUESTIONS)
final = answers["distinctiveness"]
print(CATEGORIES[final["answer"]], final["probabilities"])
for q in ["coined", "semantic_link", "imagination", "conveys", "genus"]:
    print(q, answers[q]["answer"], answers[q]["probabilities"])

Use the question wording exactly as above. The instruction and criteria text is part of the model's input, so paraphrasing it shifts the predictions.

Sub-questions and routing

The five sub-questions follow the GRPO model's doctrine decomposition:

• If coined = Yes, classify as Fanciful.
• Else if semantic_link = No, classify as Arbitrary.
• Else if genus = Yes, classify as Generic.
• Else if conveys = Yes, classify as Descriptive.
• Else if imagination = Yes, classify as Suggestive.

The head predicts the distinctiveness score directly. The routing rule is a consistency check you can run on the sub-answers, not how the final answer is produced.

Training

• Method: Unsloth FastDecisionModel with a Clef decision head, 16-bit LoRA (r = 16, alpha = 16) on unsloth/Qwen3.5-4B, 2 epochs, learning rate 2e-4, cosine schedule, batch size 8. The head was trained at Unsloth's default head learning rate.
• Data: the same 2,100 synthetic marks as Abercrombie-GRPO, balanced across the five classes and rephrased into LegalBench form. 1,890 rows were used for training. 210 held-out rows were used for temperature calibration.
• Seeds: 3407, 42 and 1234. Seed 42 is published because, among the six-question runs, it tied for best accuracy and had the highest routing agreement and the lowest LegalBench calibration error (ECE 0.145).
• Compute: about 7 minutes per run on one Colab G4 GPU.

Acknowledgements

Huge thanks to Daniel Han and the Unsloth team for decision models in Unsloth. Going from a dataset to a calibrated decision model in an afternoon on Colab is remarkable. The decision head follows the design of Cloudflare's Clef.