ChengqianMa commited on
Commit
e3b4a82
·
verified ·
1 Parent(s): a284738

Release NativeSeg and controlled baseline checkpoints

Browse files
README.md CHANGED
@@ -1 +1,35 @@
1
- Code and checkpoints will be released before 2026.12.31
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: Qwen/Qwen3.8-27B
4
+ tags:
5
+ - medical-imaging
6
+ - computed-tomography
7
+ - 3d-segmentation
8
+ - vision-language-model
9
+ - lora
10
+ ---
11
+
12
+ # NativeSeg Pretrained Checkpoints
13
+
14
+ Official pretrained LoRA + spatial decoder + gate head checkpoints and controlled baselines for **NativeSeg: Native Vision Language Segmentation of 3D Computed Tomography via Spatial Visual Tokens and Verified Local Zoom**.
15
+
16
+ - **Code Repository:** [https://github.com/step-out/NativeSeg](https://github.com/step-out/NativeSeg)
17
+ - **Model Weights:** [https://huggingface.co/ChengqianMa/NativeSeg](https://huggingface.co/ChengqianMa/NativeSeg)
18
+ - **Base Backbone:** `Qwen/Qwen3.8-27B` (frozen vision tower and frozen lower 48 language layers)
19
+
20
+ ## NativeSeg Checkpoints
21
+
22
+ | Folder | Decoder | Trainable Params | Size (`MiB`) | SHA256 | Role |
23
+ | :--- | :---: | :---: | :---: | :--- | :--- |
24
+ | `nativeseg_stage_a/` | `five_slice` | `56.00M` | 213.75 MiB | `a36729a21fbe072a3bf55eb7bcf7d4de8750c20b782085ac6662f60566cb4242` | Stage A five-slice Round-0 readout (`86.01%` Test / `85.99%` Pediatric / `85.65%` Val / `88.72%` Val6 DSC) & consensus anchor |
25
+ | `nativeseg_stage_b/` | `five_slice` | `56.00M` | 213.75 MiB | `5a63b7b06baaacd9ac4fb74c6309f2d6ab98b3122b135ed4d16c10c889f4030c` | Stage B joint end-to-end multiround model (`89.09%` Val6 DSC in Table 3b; `85.83%` Test / `86.54%` Pediatric / `85.44%` Val single-checkpoint DSC) |
26
+ | `nativeseg_local_refiner/` | `single_slice` | `34.65M` | 132.28 MiB | `28813c0a7cb0cb8056cfb922ab4e2c826665635b3ac983b0605aa42aa6b807b4` | Verified multiscale local zoom & `<err>` gate (`AUROC = 0.957` own / `0.942` 2D nnU-Net, standalone zoom & 2D/3D specialist refinement) |
27
+ | `nativeseg_single_slice/` | `single_slice` | `34.65M` | 132.28 MiB | `8bb9c3c668d6d97ca136c4c68fec85fca238178f3f96eb229419228236db3c52` | Single-slice 2D spatial-token + skip Round 0 (`F-LMM†` controlled baseline & stage-decoupled zoom initializer) |
28
+
29
+ ## Controlled Baselines (`baselines/`)
30
+
31
+ | Folder | Architecture | Trainable Params | Size (`MiB`) | SHA256 | Role |
32
+ | :--- | :---: | :---: | :---: | :--- | :--- |
33
+ | `baselines/lisa_seg_token/` | `single_slice` (`Qwen3.8-27B`) | `34.65M` | 132.28 MiB | `433e6c9ed26e6e77d5f5a75bd11f794be4987f09e5c0e5a3afe4e2276549d0b4` | Controlled `LISA†` single `[SEG]` broadcast baseline on Qwen3.8-27B (Tables 1a, 2a, 3a, 4a) |
34
+ | `baselines/densemllm_spatial_tokens/` | `single_slice` (`Qwen3.8-27B`) | `34.65M` | 132.28 MiB | `7fd21e8a82db10803f4dcfb75c65daefd72d7cb5fa10a9c39781955a38746c87` | Controlled `DenseMLLM†` 2D spatial-tokens-only baseline on Qwen3.8-27B (Tables 1a, 2a, 3a, 4a) |
35
+ | `baselines/lrn_cnn_editor/` | `CNNEditor` (2D U-Net) | `10.05M` | 38.38 MiB | `31667710b2fbc14563eb57800fe8243e4b6ddfaa1b215b14cdf044e9728e9129` | Perturbation-trained `LRN†` CNN mask refiner baseline on TotalSegmentator (`+ LRN†` in Tables 1b, 2b, 4b) |
SHA256SUMS.txt ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ a36729a21fbe072a3bf55eb7bcf7d4de8750c20b782085ac6662f60566cb4242 nativeseg_stage_a/trainable.pt
2
+ 5a63b7b06baaacd9ac4fb74c6309f2d6ab98b3122b135ed4d16c10c889f4030c nativeseg_stage_b/trainable.pt
3
+ 28813c0a7cb0cb8056cfb922ab4e2c826665635b3ac983b0605aa42aa6b807b4 nativeseg_local_refiner/trainable.pt
4
+ 8bb9c3c668d6d97ca136c4c68fec85fca238178f3f96eb229419228236db3c52 nativeseg_single_slice/trainable.pt
5
+ 433e6c9ed26e6e77d5f5a75bd11f794be4987f09e5c0e5a3afe4e2276549d0b4 baselines/lisa_seg_token/trainable.pt
6
+ 7fd21e8a82db10803f4dcfb75c65daefd72d7cb5fa10a9c39781955a38746c87 baselines/densemllm_spatial_tokens/trainable.pt
7
+ 31667710b2fbc14563eb57800fe8243e4b6ddfaa1b215b14cdf044e9728e9129 baselines/lrn_cnn_editor/trainable.pt
baselines/densemllm_spatial_tokens/checkpoint_config.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "decoder": "single_slice",
3
+ "stage": "initial",
4
+ "step": 5000,
5
+ "last_layers": 16,
6
+ "rank": 16,
7
+ "readout": "llm",
8
+ "state_residual": 0.0,
9
+ "backbone": "Qwen3.8-27B",
10
+ "description": "Controlled DenseMLLM 2D spatial-tokens-only baseline on Qwen3.8-27B",
11
+ "trainable_parameters": 34649535,
12
+ "size_bytes": 138707467,
13
+ "sha256": "7fd21e8a82db10803f4dcfb75c65daefd72d7cb5fa10a9c39781955a38746c87",
14
+ "size_mib": 132.28
15
+ }
baselines/densemllm_spatial_tokens/trainable.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7fd21e8a82db10803f4dcfb75c65daefd72d7cb5fa10a9c39781955a38746c87
3
+ size 138707467
baselines/lisa_seg_token/checkpoint_config.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "decoder": "single_slice",
3
+ "stage": "initial",
4
+ "step": 5000,
5
+ "last_layers": 16,
6
+ "rank": 16,
7
+ "readout": "token",
8
+ "state_residual": 0.0,
9
+ "backbone": "Qwen3.8-27B",
10
+ "description": "Controlled LISA single [SEG] token broadcast baseline on Qwen3.8-27B",
11
+ "trainable_parameters": 34649535,
12
+ "size_bytes": 138707467,
13
+ "sha256": "433e6c9ed26e6e77d5f5a75bd11f794be4987f09e5c0e5a3afe4e2276549d0b4",
14
+ "size_mib": 132.28
15
+ }
baselines/lisa_seg_token/trainable.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:433e6c9ed26e6e77d5f5a75bd11f794be4987f09e5c0e5a3afe4e2276549d0b4
3
+ size 138707467
baselines/lrn_cnn_editor/checkpoint_config.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "name": "baselines/lrn_cnn_editor",
3
+ "architecture": "CNNEditor",
4
+ "widths": [
5
+ 32,
6
+ 64,
7
+ 128,
8
+ 256,
9
+ 320,
10
+ 320
11
+ ],
12
+ "steps": 30000,
13
+ "batch": 32,
14
+ "lr": 0.001,
15
+ "seed": 20261031,
16
+ "trainable_parameters": 10054529,
17
+ "size_bytes": 40248812,
18
+ "size_mib": 38.38,
19
+ "sha256": "31667710b2fbc14563eb57800fe8243e4b6ddfaa1b215b14cdf044e9728e9129"
20
+ }
baselines/lrn_cnn_editor/trainable.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:31667710b2fbc14563eb57800fe8243e4b6ddfaa1b215b14cdf044e9728e9129
3
+ size 40248812
manifest.json ADDED
@@ -0,0 +1,118 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "name": "nativeseg_stage_a",
4
+ "decoder": "five_slice",
5
+ "stage": "initial",
6
+ "step": 6000,
7
+ "world_size": 8,
8
+ "last_layers": 16,
9
+ "rank": 16,
10
+ "readout": "fused",
11
+ "state_residual": 0.0,
12
+ "backbone": "Qwen3.8-27B",
13
+ "description": "Stage A five-slice spatial-token U-Net & presence gate (<pres>) checkpoint (Round 0 & consensus anchor)",
14
+ "trainable_parameters": 55995728,
15
+ "size_bytes": 224131979,
16
+ "size_mib": 213.75,
17
+ "sha256": "a36729a21fbe072a3bf55eb7bcf7d4de8750c20b782085ac6662f60566cb4242"
18
+ },
19
+ {
20
+ "name": "nativeseg_stage_b",
21
+ "decoder": "five_slice",
22
+ "stage": "agent",
23
+ "step": 4000,
24
+ "world_size": 6,
25
+ "last_layers": 16,
26
+ "rank": 16,
27
+ "readout": "fused",
28
+ "state_residual": 1.0,
29
+ "backbone": "Qwen3.8-27B",
30
+ "description": "Stage B joint end-to-end five-slice multiround NativeSeg checkpoint (89.09% Val6 / 85.83% Test / 86.54% Pediatric / 85.44% Val DSC)",
31
+ "trainable_parameters": 55995728,
32
+ "size_bytes": 224131979,
33
+ "size_mib": 213.75,
34
+ "sha256": "5a63b7b06baaacd9ac4fb74c6309f2d6ab98b3122b135ed4d16c10c889f4030c"
35
+ },
36
+ {
37
+ "name": "nativeseg_local_refiner",
38
+ "decoder": "single_slice",
39
+ "stage": "agent",
40
+ "step": 3000,
41
+ "world_size": 6,
42
+ "last_layers": 16,
43
+ "rank": 16,
44
+ "readout": "fused",
45
+ "state_residual_standalone": 1.0,
46
+ "state_residual_refiner": 1.5,
47
+ "backbone": "Qwen3.8-27B",
48
+ "description": "Verified multiscale local zoom & error gate (<err>) checkpoint (AUROC 0.957 own / 0.942 2D nnU-Net, standalone zoom & 2D/3D specialist refiner)",
49
+ "trainable_parameters": 34649535,
50
+ "size_bytes": 138707467,
51
+ "size_mib": 132.28,
52
+ "sha256": "28813c0a7cb0cb8056cfb922ab4e2c826665635b3ac983b0605aa42aa6b807b4"
53
+ },
54
+ {
55
+ "name": "nativeseg_single_slice",
56
+ "decoder": "single_slice",
57
+ "stage": "initial",
58
+ "step": 12000,
59
+ "world_size": 8,
60
+ "last_layers": 16,
61
+ "rank": 16,
62
+ "readout": "fused",
63
+ "state_residual": 0.0,
64
+ "backbone": "Qwen3.8-27B",
65
+ "description": "Single-slice 2D spatial-token + skip Round-0 checkpoint (F-LMM† controlled baseline & stage-decoupled zoom initializer)",
66
+ "trainable_parameters": 34649535,
67
+ "size_bytes": 138707467,
68
+ "size_mib": 132.28,
69
+ "sha256": "8bb9c3c668d6d97ca136c4c68fec85fca238178f3f96eb229419228236db3c52"
70
+ },
71
+ {
72
+ "name": "baselines/lisa_seg_token",
73
+ "decoder": "single_slice",
74
+ "stage": "initial",
75
+ "step": 5000,
76
+ "world_size": 4,
77
+ "last_layers": 16,
78
+ "rank": 16,
79
+ "readout": "token",
80
+ "state_residual": 0.0,
81
+ "backbone": "Qwen3.8-27B",
82
+ "description": "Controlled LISA† single [SEG] token broadcast baseline on Qwen3.8-27B",
83
+ "trainable_parameters": 34649535,
84
+ "size_bytes": 138707467,
85
+ "size_mib": 132.28,
86
+ "sha256": "433e6c9ed26e6e77d5f5a75bd11f794be4987f09e5c0e5a3afe4e2276549d0b4"
87
+ },
88
+ {
89
+ "name": "baselines/densemllm_spatial_tokens",
90
+ "decoder": "single_slice",
91
+ "stage": "initial",
92
+ "step": 5000,
93
+ "world_size": 4,
94
+ "last_layers": 16,
95
+ "rank": 16,
96
+ "readout": "llm",
97
+ "state_residual": 0.0,
98
+ "backbone": "Qwen3.8-27B",
99
+ "description": "Controlled DenseMLLM† 2D spatial-tokens-only baseline on Qwen3.8-27B",
100
+ "trainable_parameters": 34649535,
101
+ "size_bytes": 138707467,
102
+ "size_mib": 132.28,
103
+ "sha256": "7fd21e8a82db10803f4dcfb75c65daefd72d7cb5fa10a9c39781955a38746c87"
104
+ },
105
+ {
106
+ "name": "baselines/lrn_cnn_editor",
107
+ "architecture": "CNNEditor",
108
+ "stage": "refiner_baseline",
109
+ "step": 30000,
110
+ "world_size": 1,
111
+ "backbone": "2D-Residual-UNet",
112
+ "description": "Perturbation-trained LRN† CNN mask refiner baseline on TotalSegmentator (reproduces + LRN† in Tables 1b, 2b, 4b)",
113
+ "trainable_parameters": 10054529,
114
+ "size_bytes": 40248812,
115
+ "size_mib": 38.38,
116
+ "sha256": "31667710b2fbc14563eb57800fe8243e4b6ddfaa1b215b14cdf044e9728e9129"
117
+ }
118
+ ]
nativeseg_local_refiner/checkpoint_config.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "decoder": "single_slice",
3
+ "stage": "agent",
4
+ "step": 3000,
5
+ "last_layers": 16,
6
+ "rank": 16,
7
+ "readout": "fused",
8
+ "state_residual_standalone": 1.0,
9
+ "state_residual_refiner": 1.5,
10
+ "backbone": "Qwen3.8-27B",
11
+ "description": "Verified multiscale local zoom & error gate (<err>) checkpoint (standalone zoom, 2D/3D specialist refinement, and Figure 4)",
12
+ "trainable_parameters": 34649535,
13
+ "size_bytes": 138707467,
14
+ "sha256": "28813c0a7cb0cb8056cfb922ab4e2c826665635b3ac983b0605aa42aa6b807b4",
15
+ "size_mib": 132.28
16
+ }
nativeseg_local_refiner/trainable.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:28813c0a7cb0cb8056cfb922ab4e2c826665635b3ac983b0605aa42aa6b807b4
3
+ size 138707467
nativeseg_single_slice/checkpoint_config.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "decoder": "single_slice",
3
+ "stage": "initial",
4
+ "step": 12000,
5
+ "last_layers": 16,
6
+ "rank": 16,
7
+ "readout": "fused",
8
+ "state_residual": 0.0,
9
+ "backbone": "Qwen3.8-27B",
10
+ "description": "Single-slice 2D spatial-token + skip Round-0 checkpoint (F-LMM controlled baseline & stage-decoupled zoom initializer)",
11
+ "trainable_parameters": 34649535,
12
+ "size_bytes": 138707467,
13
+ "sha256": "8bb9c3c668d6d97ca136c4c68fec85fca238178f3f96eb229419228236db3c52",
14
+ "size_mib": 132.28
15
+ }
nativeseg_single_slice/trainable.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8bb9c3c668d6d97ca136c4c68fec85fca238178f3f96eb229419228236db3c52
3
+ size 138707467
nativeseg_stage_a/checkpoint_config.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "decoder": "five_slice",
3
+ "stage": "initial",
4
+ "step": 6000,
5
+ "last_layers": 16,
6
+ "rank": 16,
7
+ "readout": "fused",
8
+ "state_residual": 0.0,
9
+ "backbone": "Qwen3.8-27B",
10
+ "description": "Stage A five-slice spatial-token U-Net & presence gate (<pres>) checkpoint (Round 0 & consensus anchor)",
11
+ "trainable_parameters": 55995728,
12
+ "size_bytes": 224131979,
13
+ "sha256": "a36729a21fbe072a3bf55eb7bcf7d4de8750c20b782085ac6662f60566cb4242",
14
+ "size_mib": 213.75
15
+ }
nativeseg_stage_a/trainable.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a36729a21fbe072a3bf55eb7bcf7d4de8750c20b782085ac6662f60566cb4242
3
+ size 224131979
nativeseg_stage_b/checkpoint_config.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "decoder": "five_slice",
3
+ "stage": "agent",
4
+ "step": 4000,
5
+ "last_layers": 16,
6
+ "rank": 16,
7
+ "readout": "fused",
8
+ "state_residual": 1.0,
9
+ "backbone": "Qwen3.8-27B",
10
+ "description": "Stage B joint end-to-end five-slice multiround NativeSeg checkpoint (single-checkpoint inference & Table 3b 89.09% DSC)",
11
+ "trainable_parameters": 55995728,
12
+ "size_bytes": 224131979,
13
+ "sha256": "5a63b7b06baaacd9ac4fb74c6309f2d6ab98b3122b135ed4d16c10c889f4030c",
14
+ "size_mib": 213.75
15
+ }
nativeseg_stage_b/trainable.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5a63b7b06baaacd9ac4fb74c6309f2d6ab98b3122b135ed4d16c10c889f4030c
3
+ size 224131979