| seed: 2048 | |
| cudnn_deterministic: false | |
| train_data_jsons: | |
| - data_reason.ALL.json | |
| batch_scale: 4000 | |
| text_batch_scale: 3000 | |
| max_length: 1024 | |
| min_length: 10 | |
| n_worker: 4 | |
| local_rank: 0 | |
| minibatch_debug: -1 | |
| audio_semantic_card: 8206 | |
| audio_reason_card: 4110 | |
| local_model: meta-llama/Llama-3.2-300M | |
| parallel_number: 9 | |
| reason_pad_token: 4097 | |
| semantic_pad_token: 8193 | |
| llm_pretrained_model: understand_v3/ep1.checkpoint | |
| llm_name: meta-llama/Llama-3.2-3B | |
| text_tokenizer_path: tools/tokenizer/Text2ID/llama3_2_tokenizer | |
| semantic_eos: 8194 | |
| semantic_bos: 8195 | |
| reason_bos: 4098 | |
| reason_eos: 4099 | |
| text_pad_token: 0 | |
| exp_dir: generation_stage2 | |
| print_freq: 100 | |
| save_interval: 5000 | |
| training_stage: 2 | |
| resume: null | |
| prompt_token_path: MusicLLM2/prompts/task_prompt.pt | |
| audio_embeddings_path: extracted_params/audio_embeddings.pth | |
| audio_understanding_expert_path: extracted_params/audio_expert.pth | |
| n_epoch: 2 | |
| grad_accum: 1 | |
| learning_rate: 0.0002 | |
| grad_clip: 1.0 | |
| warmup_steps: 10000 | |
| schedule: cosine | |
| weight_decay: 0.05 | |
| beta1: 0.9 | |
| beta2: 0.95 | |
| data_parallel: fsdp | |
| mixed_precision: bf16 | |
| grad_precision: bf16 | |
| activation_checkpointing: true | |
| no_wandb: true | |
| audio_tokenizer: reasoningCodec | |
| text_tokenizer: llama3-3B | |
| rank: 0 | |