VisualEars XLS-R 300M target269h step7000

Best saved checkpoint from xlsr_visualears_cache_stream257h_b32_nogc_20260607T182456Z.

Training crashed after step ~7200 from an NCCL watchdog timeout, but best/ was preserved locally and uploaded here.

Latest completed training-log validation eval:

  • step 7000: WER 0.33466778588493573
  • step 7000: CER 0.11712760520188539
  • step 7000: eval loss 0.6059820161297403

Included artifacts:

  • model files at repo root
  • train_log.txt
  • audit/error_audit_best_sample1024/
  • audit/actionable_error_report/

Important audit note: informal-to-formal Persian drift is treated as an error for VisualEars. Examples like باشه -> باشد, می خوام -> می خواهم, and بریم -> برویم should be punished, not normalized away.

Downloads last month
7
Safetensors
Model size
0.3B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Reza2kn/visualears-xlsr-300m-target269h-step7000

Finetuned
(894)
this model