config.json

{
  "architectures": [
    "RLDX"
  ],
  "attn_implementation": null,
  "backbone_model_type": "vtc_qwen3_vl",
  "backbone_trainable_params_fp32": true,
  "color_jitter_params": {
    "brightness": 0.3,
    "contrast": 0.4,
    "hue": 0.08,
    "saturation": 0.5
  },
  "conversation_image_first": false,
  "diffusion_model_cfg": {
    "action_model_max_seq_len": 512,
    "attention_head_dim": 64,
    "depth_multi_stream": 4,
    "depth_single_stream": 8,
    "dropout": 0.2,
    "final_dropout": true,
    "num_attention_heads": 24,
    "output_dim": 1024,
    "positional_embeddings": "rope_sa_only",
    "pre_norm": "layer_norm",
    "qk_norm": "rms_norm",
    "rope_theta": 10000.0,
    "sa_dim": 1536,
    "set_triple_stream_for_mq": false,
    "set_triple_stream_for_state": false,
    "temb_type": "input_token",
    "use_swiglu": true,
    "vl_dim": 4096
  },
  "dtype": "bfloat16",
  "load_bf16": true,
  "memory_cfg": {
    "hidden_size": 4096,
    "intermediate_size": 16384,
    "max_position_embeddings": 32,
    "num_attention_heads": 16,
    "num_hidden_layers": 2,
    "num_key_value_heads": 16,
    "rms_norm_eps": 1e-05,
    "use_causal_attn": true,
    "use_rope": true
  },
  "memory_video_delta_indices": [
    -48,
    -32,
    -16,
    0
  ],
  "model_name": "RLWRLD/RLDX-1-VLM",
  "model_type": "RLDX-1",
  "n_cog_tokens": 64,
  "general_embodiment_train_ratio": 0,
  "qwen3_collator": true,
  "random_rotation_angle": null,
  "reproject_vision": false,
  "state_dropout_prob": 0.5,
  "transformers_version": "4.57.0",
  "tune_diffusion_model": true,
  "tune_llm": false,
  "tune_projector": true,
  "tune_top_llm_layers": 4,
  "tune_visual": false,
  "tune_vlln": true,
  "use_relative_action": true,
  "use_video": true,
  "video_length": 4
}