arxiv_categories_recover_c161: arXiv STEM-161 three-level category classifiers

Eight classifiers, one per arXiv subject group, that assign a scientific abstract to one of 161 categories of a three-level STEM taxonomy built on arXiv: 8 groups > 19 archives > 161 categories. Labels read GROUP_archive_category, for example PHYS_hep-ph_hep-ph.QCD, COMP_cs_cs.LG, MATH_math_math.AG.

The taxonomy respects arXiv: every arXiv category is its own category in its own group (nothing merged, nothing moved). Fourteen categories are added where arXiv has no subject classes: hep-ph (4), hep-th (3), gr-qc (3) and quant-ph (4). The full definition is in categories.json / categories.tsv; the design and its alignment with university disciplines and with CIP, OECD FORD, ANZSRC and ISCED-F are in taxonomic_align.md.

Model

The paper's group (level 1) selects the model; the model predicts the group's labels from the abstract.

• Embedding: google/embeddinggemma-300m, document mode, abstract only, normalized (not included here: download it under the Gemma terms of use).
• Head per group: LayerNorm(768) > Linear(768, 2048) > GELU > Dropout(0.15) > Linear(2048, n), softmax.
• fastText per group on the abstract text, combined log-linearly: log P = log P_head + w log P_fasttext.
• Training: per category at most 3,000 arXiv abstracts labelled with the author's primary category (for the 14 added categories, papers whose title matched exactly one category), Kaggle arXiv metadata snapshot of 2026-10-10.

Accuracy

Top-1 on held-out papers (at most 300 per category, so small categories weigh as much as large ones):

Group	Code	Categories	Training papers	EmbeddingGemma head	fastText	Ensemble
Physics	PHYS	60	161,093	72.0%	64.6%	73.3%
Computer Science	COMP	39	89,156	71.2%	64.8%	71.7%
Mathematics	MATH	31	85,753	69.8%	67.7%	73.4%
Quantitative Biology	QBIO	10	14,558	72.8%	66.6%	74.1%
Quantitative Finance	QFIN	9	6,198	59.1%	55.1%	60.1%
Statistics	STAT	5	11,353	74.5%	71.0%	76.1%
Electrical Engineering and Systems Science	EESS	4	12,000	91.5%	91.5%	92.1%
Economics	ECON	3	5,843	83.9%	83.3%	85.3%

For the 14 added categories, agreement with an LLM judge (gpt-6-astra) on 100 random papers per archive is about 72%; these fields overlap in practice, so part of the disagreement is ambiguity.

Usage

from inference import STEM161
m = STEM161()                       # downloads google/embeddinggemma-300m, or pass embedding_model="/local/path"
m.predict("Physics", "We study gravitational waves from binary black hole mergers observed by LIGO and Virgo.")
# [{'label': 'PHYS_gr-qc_gr-qc.GW', 'category': 'gr-qc.GW', 'name': 'Gravitational Waves', 'score': 0.96}, ...]

pip install -r requirements.txt. Groups: Computer Science, Economics, Electrical Engineering and Systems Science, Mathematics, Physics, Quantitative Biology, Quantitative Finance, Statistics. With fastText 0.9 builds under NumPy 2, predict one text at a time (batch prediction returns wrong probabilities); inference.py does so.

Labels per group

• Computer Science (COMP): COMP_cs_cs.AI, COMP_cs_cs.LG, COMP_cs_cs.MA, COMP_cs_cs.NE, COMP_cs_cs.CE, COMP_cs_cs.SY ...
• Economics (ECON): ECON_econ_econ.GN, ECON_econ_econ.EM, ECON_econ_econ.TH
• Electrical Engineering and Systems Science (EESS): EESS_eess_eess.AS, EESS_eess_eess.SY, EESS_eess_eess.IV, EESS_eess_eess.SP
• Mathematics (MATH): MATH_math_math.CT, MATH_math_math.AC, MATH_math_math.GR, MATH_math_math.KT, MATH_math_math.NT, MATH_math_math.QA ...
• Physics (PHYS): PHYS_physics_physics.atom-ph, PHYS_physics_physics.chem-ph, PHYS_physics_physics.atm-clus, PHYS_physics_physics.optics, PHYS_physics_physics.acc-ph, PHYS_physics_physics.app-ph ...
• Quantitative Biology (QBIO): QBIO_q-bio_q-bio.BM, QBIO_q-bio_q-bio.CB, QBIO_q-bio_q-bio.PE, QBIO_q-bio_q-bio.GN, QBIO_q-bio_q-bio.NC, QBIO_q-bio_q-bio.TO ...
• Quantitative Finance (QFIN): QFIN_q-fin_q-fin.CP, QFIN_q-fin_q-fin.GN, QFIN_q-fin_q-fin.MF, QFIN_q-fin_q-fin.PM, QFIN_q-fin_q-fin.PR, QFIN_q-fin_q-fin.EC ...
• Statistics (STAT): STAT_stat_stat.ME, STAT_stat_stat.AP, STAT_stat_stat.CO, STAT_stat_stat.ML, STAT_stat_stat.OT

Files

File	Content
config.json	groups, labels, file paths, ensemble weights
categories.json, categories.tsv	the 161 categories: label, group, archive, category, name, discipline, source
taxonomic_align.md	taxonomy design and alignment
heads/*.safetensors	the 8 MLP heads (fp16)
fasttext/*.bin	the 8 fastText models
benchmark.json	test results per group
inference.py, example.py, requirements.txt	usage

Limitations

• Trained on arXiv abstracts only; other text types (full papers, web pages) may need adaptation.
• Scores are not calibrated: the fastText part is overconfident.
• arXiv's own categories overlap; a paper's primary category is often a judgement call.