UserNeeds2K (FP16)

This standalone variant stores all weights in FP16. EmbeddingGemma computes in BF16 for numerical stability, while the classifier computes in FP16.

An extreme multi-label classifier for 2,000 categories (Categoryv2-2000, trained on userneeds_samples_v3 and 2,000 encyclopedic domain overviews). Text is embedded with the bundled google_embeddinggemma-300m model; a 2,048-unit dynamic-negative MLP (userneeds2k_dynamic_negative_mlp) produces independent category scores. The package runs without downloading a separate embedding model.

Usage

pip install -r requirements.txt
python example.py

Example output:

0.6842  10004/20035/30200  Beauty & Fitness / Fitness / Fitness Instruction & Personal Training
0.5914  10004/20035/30203  Beauty & Fitness / Fitness / General Fitness
0.4418  10004/20035/31191  Beauty & Fitness / Fitness / Home Workouts, Bodyweight & Calisthenics
0.2875  10022/20214/30775  Reference / General Reference / How-To, DIY & Expert Content
0.1943  10004/20035/30202  Beauty & Fitness / Fitness / High Intensity Interval Training

Supported language keys

Key	Language
cn	Chinese
en	English
de	German
sp	Spanish
fr	French
jp	Japanese

Performance

Test set	nDCG@5	Precision@5	Recall@5	Top-1 hit	Tail recall@5
English	0.803369	0.562866	0.758034	0.890662	0.584744
Chinese	0.756571	0.527800	0.716975	0.857000	0.503401

On an H100, this optimized FP16-storage variant processed 811.41 examples per second with 6952.34 MB peak GPU memory and achieved 0.756280 nDCG@5 on the 1,000-example precision comparison set.

No training records, translated samples, document embeddings, targets, split files, or training-derived document centroids are included in this package.