# ChunkFormer fine-tuned checkpoint

This repo contains:

- `pytorch_model.bin`: best model weights for inference / `ChunkFormerModel.from_pretrained`.
- `resume_checkpoint/training_checkpoint.pt`: latest full checkpoint for resuming training.
- `resume_checkpoint/model_state_dict.pt`: latest model state dict only.
- `resume_checkpoint/optimizer.pt`: latest optimizer state, if available.
- `resume_checkpoint/scheduler.pt`: latest scheduler state, if available.
- `resume_checkpoint/scaler.pt`: latest AMP GradScaler state, if available.
- `best_checkpoints.json`: validation loss metadata.
- `history.json`: training history.

Latest trained epoch: 2
Latest train_loss: 178.68146068524644
Latest val_loss: 155.45725427576014

Best epoch: 2
Best train_loss: 178.68146068524644
Best val_loss: 155.45725427576014

Base model: khanhld/chunkformer-ctc-large-vie