GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
milanakdj
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
milanakdj
Model Tree
openai/whisper-large-v3
Input Modalities
Output Modalities
Supported Functionality
Scored with whisper-eval.py on a held-out test split the model never saw
during training (same dataset slice, same seed 42, same 80/10/10 carve).
| Metric | Value |
|---|---|
| Test WER | 11.45% |
| Test CER | 3.08% |
| Test examples scored | 2000 of 17700 |
| Best per-epoch eval WER | 17.62% |
| Final train loss | 0.01 |
| Steps trained | 26550 |
| Epochs completed | 3.00 |
Every checkpoint scored on the same test clips, best first.
| Checkpoint | WER | CER |
|---|---|---|
checkpoint-26550 | 11.45% | 3.08% |
checkpoint-17700 | 13.40% | 3.54% |
The weights in this repo are the best row.
| Parameter | Value |
|---|---|
| Base model | openai/whisper-large-v3 |
| Language / task | nepali / transcribe |
| Epochs | 3.0 |
| Learning rate | 5e-06 |
| LR schedule | linear with 5% warmup ratio |
| Per-device train batch | 1 |
| Grad accumulation steps | 16 |
| Parameter | Value |
|---|---|
| Dataset | lilgoose7777/slr-combined-nepali-tts2 |
| Rows requested | 177000 |
| Split | 80% / 10% / 10% (train/val/test) |
| Train / val / test examples | 141600 / 17700 / 17700 |
| Audio sampling rate | 16 kHz mono |
| Checkpoint backups | milanakdj/whisper-large-v3-nepali-checkpoints |
The training corpus is clean single-speaker studio audio, so expect degraded accuracy on noisy real-world recordings with background noise, multiple speakers, or strong accents.
python
import torchfrom transformers import WhisperForConditionalGeneration, WhisperProcessor model_id = "milanakdj/whisper-large-v3-nepali-final"device = "cuda" if torch.cuda.is_available() else "cpu" processor = WhisperProcessor.from_pretrained(model_id, language="nepali", task="transcribe")model = WhisperForConditionalGeneration.from_pretrained(model_id).to(device) # audio: 1-D float32 numpy array at 16kHzinputs = processor(audio, sampling_rate=16000, return_tensors="pt").to(device)with torch.inference_mode(): ids = model.generate(inputs.input_features, max_new_tokens=225) print(processor.batch_decode(ids, skip_special_tokens=True)[0])| Effective batch size | 16 |
| Precision | bf16 |
| Gradient checkpointing | True |
| Optimizer | OptimizerNames.ADAFACTOR |
| Weight decay | 0.0 |
| Max grad norm | 1.0 |
| Generation max length | 96 |
| Eval / save strategy | per epoch (best model by WER kept) |
| Seed | 42 |