This repository provides a physically compressed and LoRA-healed checkpoint of google-bert/bert-base-uncased fine-tuned on GLUE SST-2, compressed using the svd_llm_compressor (SVDLLMCompressor) engine from the NeuralOps Library at Pareto-selected rank k=32 (3.36x compression ratio, -70.2% parameter reduction).
SVDLLMCompressor: SVD-LLM (ICLR 2025) Compressor with Cholesky activation whitening, sequential LoRA recovery fine-tuning, and fixed rank parameter ratios.
| Property | Specification |
|---|---|
Registry Key (COMPRESSOR_REGISTRY) |
svd_llm_compressor |
| NeuralOps Class & Family | SVDLLMCompressor (low_rank) |
| Source Implementation | neuralops/compression/low_rank/svd_llm_compressor.py |
| Base Architecture | google-bert/bert-base-uncased (109,483,778 parameters, 12 layers, $D=768$) |
Showcase Rank (k) |
32 (LoRA Recovery Applied: True) |
| Compressed Parameter Count | 32,585,474 (-70.24% reduction, 3.360x physical compression ratio) |
| Teacher Baseline SST-2 Acc | 91.63% |
| Showcase Final SST-2 Acc | 68.81% (Zero-Shot: 50.00%, Parity Delta: -22.82%) |
k=256, 128, 64, 32)Evaluated physically on NVIDIA GeForce RTX 3090 (24GB) on the GLUE SST-2 validation split (872 samples). Whenever zero-shot compression drops accuracy by > 2.0% alongside > 1.10x reduction, NeuralOps applies targeted r=8 LoRA healing (50 steps, lr=2e-4, 45s budget):
Rank (k) |
Physical Ratio | Compressed Params (Reduction) | Zero-Shot Acc | Healed Acc | Parity vs Teacher (91.63%) |
|---|---|---|---|---|---|
256 |
1.56x |
70,260,482 (-35.8%) |
59.52% |
53.56% (LoRA) |
-38.07% |
128 |
2.25x |
48,731,906 (-55.5%) |
52.41% |
50.92% (LoRA) |
-40.71% |
64 |
2.88x |
37,967,618 (-65.3%) |
51.38% |
50.92% (LoRA) |
-40.71% |
32 ⭐ |
3.36x |
32,585,474 (-70.2%) |
50.00% |
68.81% (LoRA) |
-22.82% |
(⭐ indicates the serialized checkpoint packaged in this Hugging Face repository.)
pytorch_model.bin: Full serialized compressed BERT state dictionary (with low-rank factors / quantization buffers / LoRA healing adapters baked in).config.json: Model configuration tagged with compression_method: "svd_llm_compressor" and compression_rank: 32.compression_summary.json: Complete machine-readable telemetry across all evaluated ranks [256, 128, 64, 32].bert-base-uncased vocabulary (vocab.txt, tokenizer.json, tokenizer_config.json).import torch
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer, BertForSequenceClassification
from neuralops.core.compression_pipeline import compress_model
repo_id = "tahamajs/bert-comp-svd-llm-compressor-sst2"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
# Reconstruct compressed NeuralOps topology and load serialized compressed + healed weights
model = BertForSequenceClassification.from_pretrained("google-bert/bert-base-uncased", num_labels=2)
model = compress_model(model, {"compressors": [{"method": "svd_llm_compressor", "rank": 32}]})
ckpt_path = hf_hub_download(repo_id=repo_id, filename="pytorch_model.bin")
state_dict = torch.load(ckpt_path, map_location="cpu")
model.load_state_dict(state_dict, strict=False)
model.eval()
inputs = tokenizer("NeuralOps achieves strong compression without sacrificing downstream accuracy.", return_tensors="pt")
with torch.no_grad():
logits = model(**inputs).logits
pred = logits.argmax(dim=-1).item()
print("Predicted SST-2 Sentiment (0=Negative, 1=Positive):", pred)
import torch
from neuralops.hub import load_surgeried_model
# Load model and tokenizer/processor in one line directly from Hugging Face Hub
model, tokenizer = load_surgeried_model("tahamajs/bert-comp-svd-llm-compressor-sst2")
model.eval()
print(f"Successfully loaded tahamajs/bert-comp-svd-llm-compressor-sst2 with NeuralOps modular surgical weights!")
@software{neuralops2026compressorzoo,
title = {NeuralOps: Multi-Rank Compressed BERT-SST2 Zoo with Pareto LoRA Recovery},
author = {Majlesi, Taha},
year = {2026},
url = {https://github.com/tahamajs/NeuralOps}
}
google-bert/bert-base-uncased