GRT-30M-NER (MatterWaveAI/grt-30m-ner)

GRT-30M-NER (GRTForTokenClassification) is a 100% Native Rotational Named Entity Recognition (NER) Model (29,307,847 parameters, zero nn.Linear layers, zero nn.Conv1d layers) built on MatterWaveAI/grt-30m for CoNLL-2003 9-class BIO entity tagging (O, B-PER, I-PER, B-ORG, I-ORG, B-LOC, I-LOC, B-MISC, I-MISC).

Each token's final GRTPearlNorm representation is L2-normalized onto the unit hypersphere and scored against 9 learned planar-pair phase-angle prototypes (phase_angles of shape [9, 128]) and softplus radial gains (pair_scales of shape [9, 128]) inside GRTCosineClassificationHead (classifier), calibrated by output_scale and per-class bias.

---

8-Way Classical Geometric Optimizer Ablation (GRTForTokenClassification)

All 8 classical geometric optimizers in transformers.models.grt.optimization_grt were benchmarked under an identical 200-step adaptation protocol (seed = 42, batch_size = 24, backbone_lr = 1.6e-4) unfreezing the top rotational transformer block blocks[-1] (811,651 parameters), final_norm (128 parameters), and classifier (2,350 parameters) for 814,257 total trainable parameters (0 nn.Linear):

Rank	Optimizer (optimization_grt.py)	Final Train Loss	Val NLL (val_loss)	Token Acc (%)	Entity Token F1 (%)	Entity Span F1 (%)	Span Precision (%)	Span Recall (%)	Wall Time (200 Steps)
1	hessian_pair_adam (HessianPairAdam (Sophia-Torus Cauchy-Riemann Curvature))	0.0584	0.1851	95.38%	91.29%	69.16%	62.03%	78.12%	24.74s
2	polar_natural_adam (PolarNaturalAdam (Riemannian Polar Metric + Cauchy-Riemann Hessian))	0.0585	0.1854	95.38%	91.29%	69.16%	62.03%	78.12%	23.07s
3	pair_adam (PairAdam (Factored Toroidal Baseline))	0.0587	0.1854	95.32%	91.19%	69.16%	62.03%	78.12%	25.32s
4	gyric_adam (GyricAdam (Poinsot-Lande Gyroscopic Projection + Thomas Precession))	0.0592	0.1863	95.32%	91.19%	69.16%	62.03%	78.12%	239.15s
5	symplectic_adam (SymplecticAdam (Gromov-Liouville Non-Squeezing Floor + Casimir Cavity))	0.0591	0.1863	95.27%	91.08%	68.78%	61.63%	77.81%	241.02s
6	huygens_symplectic_adam (HuygensSymplecticAdam (Two-Stage Huygens -> Symplectic Curriculum))	0.0591	0.1863	95.27%	91.08%	68.78%	61.63%	77.81%	218.89s
7	cyclide_adam (CyclideAdam (Dupin Cyclide Conic Metric + Deviatoric Twist))	0.0585	0.1854	95.22%	90.98%	68.41%	61.23%	77.50%	75.52s
8	huygens_adam (HuygensAdam (Circular Order Parameter R_p Splay Torque))	0.0584	0.1853	95.22%	90.98%	68.04%	60.84%	77.19%	189.88s
—	Step-0 Initialized Rotational Head (Before Adaptation)	—	0.9752	64.88%	51.84%	20.11%	14.44%	33.12%	—

---

Quickstart (transformers)

import torch
from transformers import AutoTokenizer, GRTForTokenClassification

model_id = "MatterWaveAI/grt-30m-ner"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = GRTForTokenClassification.from_pretrained(model_id)

text = "Lead researcher Elena Rostova from MatterWave AI presented the Apache License 2.0 in San Francisco yesterday."
inputs = tokenizer(text, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
pred_ids = logits.argmax(dim=-1)[0].tolist()
for tok_id, label_id in zip(inputs.input_ids[0].tolist(), pred_ids):
    print(f"{tokenizer.decode([tok_id]):18s} -> {model.config.id2label[label_id]}")

---

License

Apache 2.0