DoodDood/abercrombie-grpo

Abercrombie GRPO — Training Set

A 2,100-row synthetic training corpus for teaching a small language model to place a trademark on the Abercrombie distinctiveness spectrum (Generic → Descriptive → Suggestive → Arbitrary → Fanciful), by decomposing the judgment into five fixed yes/no doctrinal questions and applying a deterministic routing rule.

This is the training data behind DoodDood/abercrombie-grpo (a LoRA adapter, r=16, on Qwen3.5-4B). GRPO against the environment that consumes this corpus lifted LegalBench abercrombie from 40.0% → 68.4% (+28.4), and halved the mean ordinal distance (1.09 → 0.53) — attacking the base model's severe Descriptive bias (Q4=Yes for nearly everything; 0% on Arbitrary and Fanciful at baseline).

• Prime Intellect environment: smolclaims/abercrombie — dashboard
• Trained model: DoodDood/abercrombie-grpo
• Eval set: not redistributed here — the eval is the LegalBench abercrombie test split, available at nguha/legalbench. See Evaluation.

⚠️ Synthetic + model-generated. Marks, goods/services, and industries are invented (see Provenance). This dataset is for training/evaluation of legal-reasoning models — it is not legal advice and does not represent real trademark registrations or a complete statement of trademark-distinctiveness doctrine.

The task

Single-turn classification. Given a mark-and-goods phrase in the form The mark "X" for Y., the model answers five yes/no questions and a final label, in exactly this format:

Q1: [Yes/No]
Q2: [Yes/No]
Q3: [Yes/No]
Q4: [Yes/No]
Q5: [Yes/No]
FINAL_CLASSIFICATION: [Generic/Descriptive/Suggestive/Arbitrary/Fanciful]

The five questions:

• Q1 — Coined Term: Is the mark an invented/coined word?
• Q2 — Semantic Relationship: Does the mark have any semantic relationship to the goods?
• Q3 — Imagination: Does it take imagination to connect the mark to the goods?
• Q4 — Immediate Conveyance: Does the mark immediately convey a characteristic of the goods?
• Q5 — Genus: Is the mark the common/genus name for the goods?

Routing rule (frozen, deterministic):

Q1 == Yes   -> Fanciful
elif Q2 == No  -> Arbitrary
elif Q5 == Yes -> Generic
elif Q4 == Yes -> Descriptive
elif Q3 == Yes -> Suggestive

Data fields

Each line in abercrombie_train.jsonl is one JSON object:

Field	Type	Description
prompt	str	The mark-and-goods phrase, The mark "X" for Y. (user turn).
classification	str	Ground-truth label: one of Generic, Descriptive, Suggestive, Arbitrary, Fanciful.
decisive_q	str	The dispositive sub-question for this label: Q1–Q5.
decisive_q_answer	str	Correct answer to the decisive question.
industry	str	Synthetic industry / product category.
id	str	Unique row id, {label}-{NNNN} (see id scheme).

Composition

2,100 rows, roughly balanced (~420 per class) to match the 5-class ordinal target.

Class	Count	id prefix
Generic	420	generic
Descriptive	428	descriptive
Suggestive	406	suggestive
Arbitrary	426	arbitrary
Fanciful	420	fanciful
Total	2,100

The label → decisive-Q mapping

Unlike a joint mapping, the decisive question here is a function of the true label alone — it is the single sub-question that the routing rule keys on to reach that label:

Label	Decisive Q	Decisive answer
Fanciful	Q1	Yes
Arbitrary	Q2	No
Generic	Q5	Yes
Descriptive	Q4	Yes
Suggestive	Q3	Yes

Non-decisive questions are scaffolding — they are generated for format coherence but are not the graded dispositive element for that row.

id scheme

Row ids are unique and zero-padded per class prefix: generic-0001 … generic-0420, descriptive-0001 … descriptive-0428, suggestive-0001 … suggestive-0406, arbitrary-0001 … arbitrary-0426, fanciful-0001 … fanciful-0420.

Provenance and generation

• Generated from a canonical Mark: … / Goods/Services: … source and reformatted to the LegalBench surface form (The mark "X" for Y.) via reformat_to_legalbench.py.
• Surface-form match: phrased to match the LegalBench test surface form exactly, including straight quotes (the test set uses straight quotes in 92/95 rows; training matches that), so train and eval share the same surface form.
• Contamination control: marks are invented and the generator blacklist excludes every LegalBench test mark, so there is no train/test contamination.

Evaluation

The real target metric is the LegalBench abercrombie 95-row test set (19 per class), which is not included in this repo to respect LegalBench's licensing. Load it from the source instead:

from datasets import load_dataset
lb = load_dataset("nguha/legalbench", "abercrombie")

Headline result (non-thinking, greedy; base Qwen3.5-4B → +Abercrombie-GRPO LoRA):

Category	Base	+GRPO	Δ
Generic	89%	100%	+11
Descriptive	100%	74%	-26
Suggestive	5%	26%	+21
Arbitrary	0%	47%	+47
Fanciful	5%	95%	+90
Overall	40.0%	68.4%	+28.4

Mean ordinal distance: 1.09 → 0.53 (halved). ~300 GRPO steps, ~$12 of compute. The Descriptive regression is the expected cost of unwinding the base model's Descriptive-everything bias; the net is a +28.4 overall lift with the mispredictions pulled closer on the spectrum. See the model card for the frozen system prompt and the enable_thinking=False requirement.

Usage

from datasets import load_dataset

ds = load_dataset("DoodDood/abercrombie-grpo", split="train")
print(ds[0]["prompt"])
print(ds[0]["classification"], ds[0]["decisive_q"])

License

CC-BY-4.0. Free to use with attribution. The LegalBench eval set is not included and remains under its own license at nguha/legalbench.

Citation

@misc{abercrombie_grpo_2026,
  title  = {Abercrombie GRPO — Training Set},
  author = {DoodDood},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/DoodDood/abercrombie-grpo}}
}

Method lineage — same GRPO reward shape (accuracy-dominant + decisive-element gated bonus) applied to three doctrinal shapes: TOMAGPT (hearsay 3-element conjunction, +6.4%), this (trademark distinctiveness, 5-class ordinal, +28.4%), and Personal-Jurisdiction-GRPO (binary jurisdiction with general/specific routing, +8.0%).