turn_vector_head_config.json

{
  "model": {
    "model_id": "Qwen/Qwen3-VL-2B-Instruct",
    "attn_impl": "flash_attention_2",
    "prefix": "<|im_start|>assistant\n**",
    "postfix": "**<|im_end|>",
    "shift_left": true,
    "pool_mode": "mean",
    "head_hidden_dims": [
      1024
    ],
    "head_dropout": 0.0,
    "head_layer_norm": true,
    "standardize_head_inputs": false,
    "freeze_vision_tower": true,
    "modality_specs": [
      {
        "token": "<pose>",
        "n_features": 3,
        "encoder": "planar_se2",
        "encoder_kwargs": {
          "pos_scale": 4.0,
          "hidden_dims": [
            256,
            256
          ],
          "out_norm": 0.4,
          "gain_init": 1.0,
          "n_heading_harmonics": 1,
          "use_radius": false
        },
        "column": "obs_poses",
        "transform": "pose_relative_first",
        "n_raw_features": null
      }
    ]
  },
  "loss": {
    "kind": "flow_matching",
    "huber_beta": 1.0,
    "normalize_targets": false,
    "normalizer_fit_rows": 512
  },
  "target_shape": [
    1024
  ],
  "train_content_len": 1,
  "prefix_ids": [
    151644,
    77091,
    198,
    334
  ],
  "postfix_ids": [
    334,
    151645
  ],
  "flow_head_version": 1,
  "fm_n_ticks": 20,
  "fm_context_dim": 1024,
  "fm_decoder_kwargs": {
    "d_model": 128,
    "n_layers": 4,
    "n_heads": 4,
    "dim_ff": 512,
    "dropout": 0.1,
    "n_context_tokens": 8,
    "n_dims": 3,
    "use_incoming_motion": false,
    "min_period": 0.004,
    "max_period": 4.0,
    "state_scales": [
      0.03,
      0.05
    ]
  },
  "fm_config": {
    "k_samples": 8,
    "num_inference_steps": 10,
    "metric_inference_steps": 10,
    "time_alpha": 1.5,
    "time_beta": 1.0,
    "time_scale": 0.999,
    "time_offset": 0.001,
    "stratified_time": true,
    "antithetic_noise": false,
    "action_scales": [
      0.03,
      0.03,
      0.05
    ]
  }
}