experiment/v4_config.json

{
  "model_name": "llava-hf/llava-1.5-7b-hf",
  "relation": "bathroom_toilet",
  "dtype": "float16",
  "object_keywords": [
    "toilet",
    "toilets",
    "Toilet",
    "Toilets",
    "commode",
    "lavatory",
    "latrine"
  ],
  "layers": {
    "mode": "text_only",
    "specific_layer_patterns": [],
    "top_k_layers": 0
  },
  "prompts": {
    "prompts": [
      "Describe this image.",
      "Is there a toilet in this image?"
    ],
    "sampling": "uniform"
  },
  "loss": {
    "alpha": 0.0,
    "beta": 0.0,
    "kl_clip": 10.0,
    "kl_mode": "per_token",
    "suppress_weight": 10.0,
    "suppress_margin": 10.0,
    "n_sigma": 0.5,
    "epsilon": 0.05,
    "lambda_mix": 0.7,
    "smooth_tau": 0.01,
    "adaptive_lambda": true,
    "nuclear_beta": 1e-4,
    "nuclear_every_k": 10,
    "depth_weighted_nuclear": true
  },
  "lora": {
    "lora_r": 16,
    "lora_alpha": 32,
    "lora_dropout": 0.05,
    "lora_target_modules": "all-linear"
  },
  "lr": 1e-5,
  "num_epochs": 5,
  "batch_size": 1,
  "grad_accumulation_steps": 4,
  "num_workers": 4,
  "csv_path": "/home/users/ntu/cong045/scratch/testing/hallucination/CC3M-Dataset/data-qwenvl32b/bathroom_toilet_labels.csv",
  "image_dir": "CC3M-Dataset/cc3m_images/train",
  "max_train_samples": null,
  "upsample_non_scene_object": 3,
  "output_dir": "./step3_v4_outputs",
  "save_every_n_steps": 100
}