| exp_name: mme | |
| subdomain: conclusion | |
| datasets: | |
| - path-vqa | |
| - slake | |
| - vqa-rad | |
| dataset_directory: Model_Training/Dataset/train | |
| evaluation_model_priority: | |
| - gpt4 | |
| - gpt35 | |
| n_epochs: 1 | |
| n_examples: null | |
| n_eval_examples: 10 | |
| batch_size: 128 | |
| gradient_accumulation_steps: 64 | |
| eval_batch_size: 2 | |
| max_length: 3520 | |
| max_prompt_length: 2924 | |
| reward_token: Unified | |
| frozen_layers: | |
| enable: true | |
| language_model_layers: | |
| - '0' | |
| - '1' | |
| - '2' | |
| - '3' | |
| - '4' | |
| - '5' | |
| - '6' | |
| - '7' | |
| - '8' | |
| - '9' | |
| - '10' | |
| - '11' | |
| - '12' | |
| - '13' | |
| - '14' | |
| - '15' | |
| - '16' | |
| - '17' | |
| - '18' | |
| - '19' | |
| - '20' | |
| - '21' | |
| - '22' | |
| - '23' | |
| vision_tower_layers: -1 | |
| debug: false | |
| fsdp_port: 46031 | |
| wandb: | |
| enabled: true | |
| entity: null | |
| project: direct-preference-optimization | |
| local_dirs: | |
| - /scr-ssd | |
| - /scr | |
| - .cache | |
| sample_during_eval: false | |
| n_eval_model_samples: 2 | |
| do_first_eval: true | |
| local_run_dir: .cache/lhe_exc/mme_2024-04-23_01-31-43_653905 | |
| lr: 1.0e-06 | |
| max_grad_norm: 10.0 | |
| trainer: FSDPTrainer | |
| optimizer: RMSprop | |
| warmup_steps: 6 | |
| activation_checkpointing: false | |
| eval_every: 14976 | |
| minimum_log_interval_secs: 1.0 | |
| seed: 0 | |
| model: | |
| name_or_path: ./LM/llava-1.5-7b-hf-medllama2 | |
| tokenizer_name_or_path: null | |
| archive: ./LM/mme/mme_conclusion/sft_only/Unified/policy.pt | |
| block_name: LlamaDecoderLayer | |
| policy_dtype: bfloat16 | |
| fsdp_policy_mp: bfloat16 | |
| reference_dtype: bfloat16 | |
| vision_encoder_block: CLIPEncoderLayer | |
| language_model_block: LlamaDecoderLayer | |
| loss: | |
| name: dpo | |
| beta: 0.1 | |
| reference_free: false | |