run_id: xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6
stage: qwen_lora_text_video_audio
backbone_id: qwen3_omni_lora
dataset_contract: xperience10m_episode_json_qa_v1
model_id: <model-cache>/Qwen__Qwen3-Omni-30B-A3B-Instruct
dataset_jsonl: results/omni_finetune/xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_dataset/dataset.jsonl
checkpoint_dir: <project>/checkpoints/xperience10m_qwen3_omni_128ep_fullsplit_fast8gpu_lora_fsdp_full_train_noval_tail_logits_fullstatesave_v6/adapter_lora
num_processes: 8
epochs: 1
learning_rate: 0.0001
lora_r: 16
lora_alpha: 32
loss_mode: answer_token_ce
loss_logit_tail_only: True