tahamajs/bert-cosformer-sst2

🧠 BERT-Base (cosformer) — Subquadratic Attention Zoo

This repository provides an authoritative pretrained and distilled checkpoint of bert-base-uncased where standard quadratic softmax attention blocks across all 12 transformer layers have been replaced with the cosformer mechanism from the NeuralOps Library.

The model was trained and evaluated on the General Language Understanding Evaluation (GLUE) benchmark suite (SST-2 task) using the NeuralOps Cached Representation Distillation Protocol (Zero Teacher VRAM allocation, offline representation alignment) physically executed on an NVIDIA GeForce RTX 3090 (24GB).

---

🔬 Architectural & Complexity Profile

• Attention Mechanism: cosformer
• Subquadratic Scaling: $O(N)$ or $O(N \log N)$ effective operational complexity versus standard quadratic $O(N^2)$ Multi-Head Attention (MHA).
• Weights Serialization: attention_only_fp16 (attention_weights.safetensors).
• Memory Footprint: ~55 MB download footprint (strictly the trained attention weights and task classification head, zero redundant base model bloat).

---

📊 Empirical Evaluation & Distillation Benchmark

Evaluated on the GLUE Benchmark SST-2 validation split (872 binary sequence classification samples):

Metric	Teacher Baseline (bert-base-uncased)	Pre-KD Student (cosformer)	Post-KD Distilled Model	Accuracy Delta ($\Delta$)	Distillation Status
GLUE SST-2 Accuracy	91.63%	91.63%	91.51%	-0.11%	Excellent (>=90%)
Validation Cross-Entropy	0.2428	Uncalibrated	0.3044	Converged	Optimal

Hardware & Distillation Telemetry (Sharif HPC Cluster)

• Physical Compute: NVIDIA GeForce RTX 3090 24GB (Ampere Architecture, Tensor Cores active)
• Training Duration: 47.9s
• Total Optimization Steps: 144 steps (Stage 1 Subspace Hidden Alignment $\rightarrow$ Stage 2 Logit KL + Task CE)
• Effective Throughput: 24617.9 tokens/sec
• Peak GPU VRAM: 13512 MB (Zero teacher VRAM overhead via offline representation caching)

---

🚀 Quick Start & One-Liner Inference

1. Standard Hugging Face Transformers (AutoModelForSequenceClassification)

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

# Instant download and execution directly from Hugging Face Hub (trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained("tahamajs/bert-cosformer-sst2", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("tahamajs/bert-cosformer-sst2")

inputs = tokenizer("NeuralOps accelerates transformer architectures with subquadratic efficiency.", return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
    prediction = outputs.logits.argmax(dim=-1).item()

print("Predicted Sentiment (0 = Negative, 1 = Positive):", prediction)

2. High-Level Modular Loading via NeuralOps Zoo API

from neuralops.hub.bert_zoo import load_pretrained_bert

# Reconstruct and load strictly the attention weights onto base BERT
model, tokenizer = load_pretrained_bert(
    attention_type="cosformer",
    task="sequence-classification",
    hf_org="tahamajs",
)

---

📦 Modular Attention-Only Weights

Following the NeuralOps parameter-efficient serialization specification, this repository stores only the specialized attention and task adapter parameters (attention_weights.safetensors in FP16, ~55 MB) rather than re-uploading the entire 450 MB frozen base BERT weights. The remote code shim (modeling_neuralops_bert.py) automatically initializes the frozen base weights from bert-base-uncased and applies the distilled attention parameters on the fly.

---

📜 Citation

If you use this model or the NeuralOps attention zoo in your research, please cite:

@software{neuralops2026bertzoo,
  title   = {NeuralOps: Pretrained BERT Subquadratic Attention Zoo on GLUE Benchmark},
  author  = {Majlesi, Taha},
  year    = {2026},
  url     = {https://github.com/tahamajs/NeuralOps}
}