Views
No views yet
1pip install huggingface_hub
2huggingface-cli download onnoboru/Qwen3.5-0.8B-FFT-LAP-UR5e-Curriculum --local-dir ./Qwen3.5-0.8B-FFT-LAP-UR5e-Curriculum1import pickle
2import torch
3from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLProcessor
4
5ckpt_dir = "./Qwen3.5-0.8B-FFT-LAP-UR5e-Curriculum"
6
7model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
8 f"{ckpt_dir}/model_final", torch_dtype=torch.bfloat16, device_map="auto",
9)
10processor = Qwen2_5_VLProcessor.from_pretrained(f"{ckpt_dir}/model_final")
11
12# Load dataset stats (required for action denormalization)
13with open(f"{ckpt_dir}/dataset_stats.pkl", "rb") as f:
14 dataset_stats = pickle.load(f)1from rv_train.train import get_pretrained_model
2
3model, cfg = get_pretrained_model("./Qwen3.5-0.8B-FFT-LAP-UR5e-Curriculum", device=0)
4model.eval()dataset_stats.pkl1import pickle
2
3with open("dataset_stats.pkl", "rb") as f:
4 stats = pickle.load(f)
5# stats contains mean/std for action dimensionsmain holds the recommended/final weights.
Earlier training-step snapshots are published as branches named step-<global_step> (e.g., step-17000, step-18000).
Load any of them by passing revision=:1# Download a specific revision
2huggingface-cli download onnoboru/Qwen3.5-0.8B-FFT-LAP-UR5e-Curriculum --revision step-18000 --local-dir ./Qwen3.5-0.8B-FFT-LAP-UR5e-Curriculum-step-18000
3
4# Or load directly via transformers
5Qwen2_5_VLForConditionalGeneration.from_pretrained(
6 "onnoboru/Qwen3.5-0.8B-FFT-LAP-UR5e-Curriculum",
7 revision="step-18000",
8 subfolder="model_final",
9)Qwen/Qwen3.5-0.8B1DATALOADER:
2 ROBOVERSE:
3 cfg_opts: IMAGE.crop_img:0.9:IMAGE.img_size:224:IMAGE.cam_list:('3p1','wrist_right1')
4 cfg_path: libs/RoboVerse/roboverse/configs/ur5e_cluttered_pick_3obj_120.yaml
5 batch_size: 16
6 num_workers: 8
7EXP:
8 AMP: true
9 DATASET: roboverse
10 EXP_ID: lap_qwen3_5_08b_fft_ur5e_cluttered_pick_3obj_120_4gpu_curriculum
11 LOSS:
12 action_weight_final: 1.0
13 curriculum_total_steps: 12000
14 think_action_curriculum: true
15 LR_SCHED: none
16 MODEL: qwen
17 OPTIMIZER: adamw
18 SEED: 0
19EXP_EXTRA:
20 no_test: true
21 no_track: true
22 no_val: true
23 save_at_steps:
24 - 10000
25 - 12000
26 - 14000
27 - 16000
28 save_ckp: 0
29 save_last_ckpt: true
30 test_eval_freq: 1
31 val_eval_freq: 1
32LR_SCHED:
33 lr_clip: 1.0e-08
34 lr_decay_factor: 0.5
35 lr_patience: 4
36MODEL:
37 QWEN:
38 action_mask_aug_per: 0.4
39 action_type: original
40 add_vision_id: true
41 attention_dropout: 0.0
42 enable_thinking: true
43 grad_checkpoint: false
44 history: 1
45 horizon: 8
46 lap_action_is_absolute: true
47 lap_emit_holds: false
48 lap_rotation_precision: 1
49 lap_sum_decimal: 1f
50 lora_config: default
51 lora_rank: 8
52 num_bins_actions: 1000
53 num_cam: 2
54 original_action_dim: 7
55 qwen_model_id: Qwen/Qwen3.5-0.8B
56 reasoning: true
57 rgb_img_size:
58 - 224
59 - 224
60 rgb_input: true
61 tiled_rgb_imgs: true
62 use_flash_attention_2: false
63 use_lora: false
64 use_qlora: false
65TRAIN:
66 clip_grad_norm: 0.0
67 l2: 1.0e-10
68 lr: 1.0e-05
69 num_epochs: 100
70 num_iters: 16000
71 save_iter_ckp: 2500
72WANDB:
73 # dir: logs
74 enable: true
75 entity: ''
76 log_interval: 50
77 mode: online
78 project: vla0
79 resume_id: ''
80 run_name: ''
81 tags: ''
82
83| File | Description |
|---|---|
model_final/model-*.safetensors | Full model weights |
model_final/config.json | Model configuration |
model_final/tokenizer.json | Tokenizer |
dataset_stats.pkl | Action normalization statistics (required for inference) |
config.yaml | Training configuration |