CoolFace
Modelpublic

Reza2kn/visualears-xlsr-300m-v2-cont-step1568-fp16

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face
0likes9downloads
Model Card

VisualEars XLS-R 300M v2 Step1568 FP16

This is the FP16 export of `Reza2kn/visualears-xlsr-300m-v2-cont-step1568` for the VisualEars Persian ASR relabeling pipeline.

The FP16 path was selected after parity checks against FP32. Lower precision compressed variants were rejected for relabeling.

Precision Decision

Accepted:

  • —float16

Rejected for relabeling:

  • —NVFP4
  • —NVFP4A16
  • —FP8
  • —FP8_DYNAMIC

FP16 Parity

269-row validation parity against FP32:

  • —Char similarity vs FP32: 0.9995259540175397
  • —CER vs FP32: 0.0009786651008025053
  • —WER vs FP32: 0.005176876617773943

Golha v2 benchmark:

ModeWERCERMean decode
FP3225.1066%7.6194%98.1 ms
FP1625.0457%7.6194%31.8 ms

Golha v2 FP16-vs-FP32:

  • —Char similarity vs FP32: 0.9994312796208531
  • —Changed normalized rows: 3 / 69

Usage

python
import torch
from transformers import AutoModelForCTC, AutoProcessor

model_id = "Reza2kn/visualears-xlsr-300m-v2-cont-step1568-fp16"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCTC.from_pretrained(model_id, dtype=torch.float16).cuda().eval()

Snapshot State

Pipeline state, parity logs, and Golha v2 comparison artifacts are archived at:

`Reza2kn/visualears-full-pipeline-state`

Snapshot path:

snapshots/2026-06-08-fp16-ready/