chenyuanxin/RT-accompanimentV2-checkpoints

RT-accompanimentV2 Checkpoints

This repository stores experimental RT-accompanimentV2 checkpoints.

These checkpoints are not packaged as standard Hugging Face Transformers from_pretrained models. They are intended to be loaded with the RT-accompanimentV2 codebase and tokenizer/model definitions from the project repository.

Contents

initdrop8p0p3/
  model.safetensors
  config.json
  run_config.json

initdrop8p0p3_melloss/
  model.safetensors
  config.json
  run_config.json

Checkpoints

initdrop8p0p3

Training setting:

• Base checkpoint: checkpoints-resume/epoch_15_0307_1858/model.safetensors
• Initial accompaniment drop: first 8 accompaniment beats are dropped with probability 0.3 per training sample
• Melody loss: disabled
• Source local checkpoint: checkpoints-teacher/music_transformer_initdrop8p0p3_0407_1138/epoch_3_0407_1740/model.safetensors

initdrop8p0p3_melloss

Training setting:

• Base checkpoint: checkpoints-resume/epoch_15_0307_1858/model.safetensors
• Initial accompaniment drop: first 8 accompaniment beats are dropped with probability 0.3 per training sample
• Melody loss: enabled
• Source local checkpoint: checkpoints-teacher/music_transformer_initdrop8p0p3_melloss_0407_2000/epoch_3_0408_0205/model.safetensors

Notes

• run_config.json records the training configuration used for each run.
• config.json records the saved model/config metadata from the selected epoch directory.
• The repository is private by default and intended for project collaboration.