earino/ecbs5200-jev-benchmark

Zero-shot classifiers, retrieved examples and trained models on 113-class CFPB complaint labels

Release accompanying Zero-Shot Classifiers, Retrieved Examples and Trained Models on a 113-Class Complaint Taxonomy (Ariño de la Rubia, 2026). Code, paper source and a number trace are in the paper's GitHub repository, https://github.com/earino/zero-shot-complaint-benchmark, which fetches the two largest files here at a pinned revision.

Everything needed to regenerate every number, table and figure in the paper without training models or calling any API. The release was verified by rebuilding it into an empty directory, installing only requirements.txt, and regenerating: results/numbers.json matched results/numbers_reference.json exactly and every generated table was byte-identical to the paper's.

Reproduce

pip install -r requirements.txt      # exact versions used to build and verify this release
python code/analysis.py --release .
python code/make_tables.py --release .

This writes results/numbers.json (compare with results/numbers_reference.json), the figures to results/figures/ and the LaTeX tables to results/tables/. The complaint texts are loaded from the public dataset determined-ai/consumer_complaints_medium (revision in MANIFEST.json) using utils/data_utils.py and the split indices in data/; they are needed for the duplicate audit, the cluster bootstrap, the label-mapping sensitivity check and the historical-label test.

Systems

Zero-shot: TypeSafe's Jev 1.13 (bare labels, form instruction, form instruction + definitions); SemIf with Qwen3.5-4B (two-round bracket, bracket + definitions, one-versus-rest, and the bracket under two further label groupings, seeds 7 and 123); GLM-5.3 via the LunaRoute gateway. Retrieval without training: GLM-5.3 with the 20 most similar labeled complaints in the prompt; an ablation that shows GLM-5.3 only the labels of those complaints, without their texts; a vote among the same 20; and the label of the single most similar complaint. Logistic regression on the fit set: TF-IDF, Jev probabilities, sentence embeddings, label-name similarity, Jev's top choice, stacked embedding probabilities, and combinations; a learning curve with hyperparameters chosen within each label budget; a 20,000-complaint subsample comparing SemIf and Jev probabilities as features. Fine-tuned: ModernBERT-base (vanilla and distilled), ModernBERT-large, Qwen3-32B + LoRA (predictions from the Hugging Face repositories listed in MANIFEST.json).

Contents

Path	What
predictions/val_predictions_all_systems.npz	For each of the 6,430 evaluation rows: true label, frequency group, SHA-256 of the complaint text, and every system's predicted label and 113-way probabilities (plus Jev's confidence). Keys in MANIFEST.json under prediction_file_keys. An unmatched generated answer is stored as label 113 (out of taxonomy).
jev_responses/*.jsonl	Raw Jev responses: chosen label, confidence, probabilities in label order, client latency, tokens, model, request id; bare labels on evaluation, train and test, and the criteria variants.
results/semif/	SemIf raw option logits per complaint, composed probabilities, manifests (SemIf commit, model revision, weight checksums, group partition) and results.
results/glm/	GLM-5.3 raw answers (zero-shot, with 20 retrieved examples, and with only the labels of those examples), including the first attempts of answers re-requested after truncation.
results/	Reference numbers, feature-control and learning-curve results, all hyperparameter-selection scores, the 20 retrieved neighbours of each evaluation complaint, latency statistics, and a token-usage ledger.
criteria/	The label-definition variants and the fixed 1,000-row training dev slice.
data/, utils/	Label list, label mapping, split indices, and the loader that applies them.
code/	Every script that produced a result (Jev, SemIf, GLM-5.3, embedding and feature controls, nearest-neighbour vote, SemIf features), the analysis, the table generator, and this release's builder.
MANIFEST.json	Model, SDK and gateway details, dataset revision, comparator repositories and revisions, and SHA-256 of every file.

Development status

Some systems' development used the evaluation split. Each system is labeled in the paper: A no use of evaluation results in development; B indirect use; C checkpoint or temperature selected on the evaluation split. Comparisons among status-A systems are the paper's confirmatory core.

Notes

• Targets are our preprocessed version of the issue the consumer selected on the CFPB form (a fixed mapping in data/label_merge_mapping.json; three entries narrow a label's meaning, and the paper reports that removing the 15 affected evaluation complaints changes no conclusion).
• 11.7% of evaluation complaints repeat another complaint in the public dataset; the paper reports results with and without near-duplicates.
• Jev latencies were measured from a single client in Vienna, Austria; SemIf ran on an Apple M4 Max laptop.
• Complaint text originates from the U.S. CFPB Consumer Complaint Database via determined-ai/consumer_complaints_medium; Jev outputs were produced with TypeSafe's API and GLM-5.3 outputs through the LunaRoute gateway. Check the upstream terms before reuse.