Censer.8 / reports /benchmark.md
KucLab's picture
Add report benchmark.md
ed82a75 verified
|
Raw History Blame Contribute Delete
1.03 kB

Censer.8 — Benchmark

Generated by scripts/benchmark.py.

Hardware

  • Device: AMD Radeon Graphics (cuda)
  • torch: 2.13.0+rocm7.1
  • Python: 3.12.12
  • OS: Linux-7.1.7-100.fc43.x86_64-x86_64-with-glibc2.43

Model

Property Value
Parameters 903,552
Checkpoint size 3.46 MB
Vocabulary 2048
Context 4096 tokens
Input limit 4000 characters
Layers / d_model / heads / d_ff 3 / 128 / 4 / 384

Inference

Metric Value
Single text latency 1.27 ms
Single text throughput 790.2 texts/s
Batch of 8 latency 1.41 ms
Batched throughput 5659.4 texts/s
Peak VRAM (inference) 84.7 MB

Accuracy

Split Rows Accuracy F1 Precision Recall
val 6585 0.8583 0.8217 0.8638 0.7835
test 6496 0.8568 0.8198 0.8622 0.7813
probe (hand-written) 97 1.0000 1.0000 1.0000 1.0000

The probe set is hand-written and absent from all training data.