UserNeeds1K (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 1,091 categories. Text is embedded with the bundled google_embeddinggemma-300m model; a 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:

0.6411  10004/20035/30203  Beauty & Fitness / Fitness / Other
0.4917  10004/20035/30200  Beauty & Fitness / Fitness / Fitness Instruction & Personal Training
0.2998  10022/20214/30775  Reference / General Reference / How-To, DIY & Expert Content
0.0853  10004/20035/30202  Beauty & Fitness / Fitness / High Intensity Interval Training
0.0628  10002/20013/30078  Arts & Entertainment / TV & Video / Online Video

Supported language keys

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

Performance

Test set	nDCG@5	Precision@5	Recall@5	Top-1 hit	Tail recall@5
English	0.858722	0.513768	0.833800	0.910088	0.629620
Chinese	0.821382	0.491000	0.792756	0.890000	0.583333

On an H100, this optimized FP16-storage variant processed 825.51 examples per second with 2,176.64 MB peak GPU memory and achieved 0.855899 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.