Configuration Parsing Warning:Invalid JSON for config file config.json
FLASH3 — ROBERTA Benchmark Result
Status: ℹ️ PASS (Efficient Attention / Zero Param Reduction)
Dataset: GLUE SST-2 (Binary)
Architecture: roberta
Module: flash3
This repository hosts empirical benchmark results and factorized checkpoints for flash3 applied to roberta on GLUE SST-2 (Binary).
Note: Weights not bundled or pruned; comprehensive profiling metrics, configurations, and verification card are preserved for research reproducibility.
Benchmark Performance Metrics
| Metric | Value |
|---|---|
| Model Family | roberta |
| Module / Method | flash3 |
| Dataset | GLUE SST-2 (Binary) |
| Validation Status | PASS_NO_COMPRESSION |
| Validation Accuracy | 62.61% |
| Pass Threshold (3*SE Margin) | 56.0% |
| F1 Score (Macro) | 0.6244 |
| F1 Score (Weighted) | 0.6502 |
| Precision | 0.6269 |
| Recall | 0.6251 |
| Validation Loss | 0.7882 |
| Compression Factor | 1.0x |
| Peak GPU VRAM | 648.38 MB |
| Inference Throughput | 15970.12 tokens/sec |
| Evaluation Time | 18.37 s |
| Parameters (Before) | 124647170 |
| Parameters (After) | 132913922 |
| Compressed Layers | 12 |
| LoRA Trainable Params | 1179648 |
| Benchmark Timestamp | 2026-09-15T14:56:09.673836 |
Architecture & Module Specification
The flash3 module is evaluated under the Hoosha AI NeuralOps SVD & Linear Attention Suite, benchmarking subquadratic attention and rank-adaptive tensor factorization against full quadratic baselines.
Framework & Verification
Developed and benchmarked using the SVD Linear Attention Framework — Universal GPU Benchmark Suite.
- Repository: tahamajs/SVD_linear_Attention
- Organization: Hooshaai
- Downloads last month
- 1