Hugging Face's logo

Sssunset
/
MedPLIB-BRISC

Image Segmentation
English
medical
brain-tumor
mri
segmentation
classification
multimodal
lora
medplib
Model card Files
xet
 Community
Sssunset commited on
Commit
d1acd9e
·
verified ·
1 Parent(s): 8dbf216

Upload MedPLIB-BRISC weights and model card

Browse files
Files changed (7) hide show
  1. .gitattributes +1 -0
  2. README.md +96 -0
  3. adapter_meta.json +355 -0
  4. assets/fig_architecture.png +3 -0
  5. assets/fig_results.png +0 -0
  6. assets/fig_stages.png +0 -0
  7. trainable.pt +3 -0
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ assets/fig_architecture.png filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ base_model: Huangxs/MedPLIB-7b-2e
4
+ language:
5
+ - en
6
+ pipeline_tag: image-segmentation
7
+ tags:
8
+ - medical
9
+ - brain-tumor
10
+ - mri
11
+ - segmentation
12
+ - classification
13
+ - multimodal
14
+ - lora
15
+ - medplib
16
+ ---
17
+
18
+ # MedPLIB-BRISC
19
+
20
+ **Joint brain tumor classification and segmentation from MRI with an adapted multimodal large language model.**
21
+
22
+ This repository holds the weights we trained on top of MedPLIB-7b-2e for the BRISC 2025 brain tumor MRI dataset. Given an MRI slice and a fixed instruction, the model answers with one of four classes and a `<SEG>` token. The four classes are glioma, meningioma, pituitary tumor and non-tumorous. The hidden state of the `<SEG>` token is decoded into the tumor mask, so the class and the mask come from the same response.
23
+
24
+ [Project page](https://huggingface.co/spaces/Sssunset/MedPLIB-BRISC) · [Code on GitHub](https://github.com/Peilin-FF/Tumor) · [Base model](https://huggingface.co/Huangxs/MedPLIB-7b-2e)
25
+
26
+ ![Model overview](assets/fig_architecture.png)
27
+
28
+ ## Files
29
+
30
+ | File | Content |
31
+ |---|---|
32
+ | `trainable.pt` | All trained weights, 286.8M parameters in total: LoRA on the language model, the image projector, the router, the `<SEG>` projector, the SAM-Med2D mask decoder and the SAM-Med2D encoder adapters |
33
+ | `adapter_meta.json` | LoRA settings and the list of trained tensors |
34
+ | `assets/` | Figures used on this page |
35
+
36
+ These weights are not a complete model. They are loaded on top of the original MedPLIB-7b-2e checkpoint, CLIP ViT-L/336 and SAM-Med2D-B with the code from our GitHub repository.
37
+
38
+ ## How to use
39
+
40
+ ```bash
41
+ git clone https://github.com/Peilin-FF/Tumor && cd Tumor
42
+ scripts/setup_env.sh # conda environment "medplib"
43
+ scripts/download_models.sh # MedPLIB-7b-2e, CLIP ViT-L/336, SAM-Med2D-B
44
+ huggingface-cli download Sssunset/MedPLIB-BRISC --local-dir checkpoints/medplib-brisc
45
+
46
+ # interactive demo
47
+ python -m medplib_bt.demo_app --adapter checkpoints/medplib-brisc --port 7860
48
+
49
+ # reproduce the test results, which also needs scripts/download_data.sh
50
+ scripts/run_infer.sh 0,1,2,3 outputs/preds/medplib-brisc --adapter checkpoints/medplib-brisc
51
+ ```
52
+
53
+ ## Training
54
+
55
+ The model is trained on the 5,000 training slices of the official BRISC 2025 split with 8 NVIDIA A100 GPUs.
56
+
57
+ | Stage | Epochs | Trained modules | Trainable parameters | Loss |
58
+ |---|---|---|---|---|
59
+ | A: Alignment | 1 | image projector, router, `<SEG>` projector, mask decoder | 43.1M, 0.37% | CE + 2 BCE + 0.5 Dice |
60
+ | B: Adaptation | 12 | Stage A modules, LoRA with rank 16 on attention and both experts in all 32 layers, SAM-Med2D encoder adapters | 286.8M, 2.44% | CE + 2 BCE + 2 Dice + IoU |
61
+
62
+ Half of the training prompts ask for the class and the mask, a quarter ask for the class only and a quarter ask for the mask only. Non-tumorous slices are trained with an empty mask. At test time, each of the four valid answers is scored by its log-likelihood and the best one is kept. The mask comes from the `<SEG>` state of that answer, and no rule removes it afterwards.
63
+
64
+ ![Two training stages](assets/fig_stages.png)
65
+
66
+ ## Results
67
+
68
+ Results on the 1,000 BRISC test slices, from one training run with seed 42:
69
+
70
+ | Metric | Value |
71
+ |---|---|
72
+ | Accuracy | 0.980 |
73
+ | Macro-F1 | 0.982 |
74
+ | Dice on the 860 tumor slices | 0.854 |
75
+ | Dice on all 1,000 slices | 0.874 |
76
+ | Masks drawn on the 140 non-tumorous slices | 0 |
77
+ | Slices where the class and the mask disagree | 0 |
78
+
79
+ Two separately trained models, EfficientNet-B0 and U-Net, draw a mask on 61.4% of the non-tumorous slices and contradict each other on 9.0% of all slices. Trained on both tasks instead of segmentation alone, the same MedPLIB model stops drawing false masks entirely. The project page has the full comparison and case studies.
80
+
81
+ ![Results](assets/fig_results.png)
82
+
83
+ ## Limitations
84
+
85
+ - The model is trained and tested on 2D slices from one public dataset. The dataset has no patient identifiers, so slices from one patient may appear in both splits.
86
+ - A dedicated U-Net still outlines tumors slightly better on average. Gliomas and small lesions are the weakest cases.
87
+ - Free text answers became very short after adaptation.
88
+ - This is a course research prototype. It is not a medical device and must not be used for diagnosis or treatment decisions.
89
+
90
+ ## Team
91
+
92
+ GP8001 Group 3, Nanyang Technological University: Simon Tong Sing Hee, Zeng Yi, Feng Peilin, Ye Xiaomeng, Lyu Muyang, Timothy Aw Bang Hao.
93
+
94
+ ## Acknowledgements
95
+
96
+ This work builds on [MedPLIB](https://arxiv.org/abs/2412.09278), [SAM-Med2D](https://arxiv.org/abs/2308.16184) and the [BRISC 2025 dataset](https://www.nature.com/articles/s41597-026-06753-y). We thank their authors for releasing code, weights and data.
adapter_meta.json ADDED
@@ -0,0 +1,355 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "n_tensors": 872,
3
+ "n_params": 286811620,
4
+ "lora": {
5
+ "peft_type": "LORA",
6
+ "auto_mapping": null,
7
+ "base_model_name_or_path": "/mnt/data/peilin/HF_MODEL/MedPLIB-7b-2e",
8
+ "revision": null,
9
+ "task_type": "CAUSAL_LM",
10
+ "inference_mode": false,
11
+ "r": 16,
12
+ "target_modules": [
13
+ "model.layers.0.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
14
+ "model.layers.0.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
15
+ "model.layers.0.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
16
+ "model.layers.0.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
17
+ "model.layers.0.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
18
+ "model.layers.0.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
19
+ "model.layers.0.self_attn.k_proj",
20
+ "model.layers.0.self_attn.o_proj",
21
+ "model.layers.0.self_attn.q_proj",
22
+ "model.layers.0.self_attn.v_proj",
23
+ "model.layers.1.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
24
+ "model.layers.1.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
25
+ "model.layers.1.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
26
+ "model.layers.1.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
27
+ "model.layers.1.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
28
+ "model.layers.1.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
29
+ "model.layers.1.self_attn.k_proj",
30
+ "model.layers.1.self_attn.o_proj",
31
+ "model.layers.1.self_attn.q_proj",
32
+ "model.layers.1.self_attn.v_proj",
33
+ "model.layers.10.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
34
+ "model.layers.10.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
35
+ "model.layers.10.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
36
+ "model.layers.10.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
37
+ "model.layers.10.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
38
+ "model.layers.10.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
39
+ "model.layers.10.self_attn.k_proj",
40
+ "model.layers.10.self_attn.o_proj",
41
+ "model.layers.10.self_attn.q_proj",
42
+ "model.layers.10.self_attn.v_proj",
43
+ "model.layers.11.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
44
+ "model.layers.11.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
45
+ "model.layers.11.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
46
+ "model.layers.11.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
47
+ "model.layers.11.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
48
+ "model.layers.11.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
49
+ "model.layers.11.self_attn.k_proj",
50
+ "model.layers.11.self_attn.o_proj",
51
+ "model.layers.11.self_attn.q_proj",
52
+ "model.layers.11.self_attn.v_proj",
53
+ "model.layers.12.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
54
+ "model.layers.12.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
55
+ "model.layers.12.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
56
+ "model.layers.12.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
57
+ "model.layers.12.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
58
+ "model.layers.12.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
59
+ "model.layers.12.self_attn.k_proj",
60
+ "model.layers.12.self_attn.o_proj",
61
+ "model.layers.12.self_attn.q_proj",
62
+ "model.layers.12.self_attn.v_proj",
63
+ "model.layers.13.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
64
+ "model.layers.13.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
65
+ "model.layers.13.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
66
+ "model.layers.13.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
67
+ "model.layers.13.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
68
+ "model.layers.13.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
69
+ "model.layers.13.self_attn.k_proj",
70
+ "model.layers.13.self_attn.o_proj",
71
+ "model.layers.13.self_attn.q_proj",
72
+ "model.layers.13.self_attn.v_proj",
73
+ "model.layers.14.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
74
+ "model.layers.14.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
75
+ "model.layers.14.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
76
+ "model.layers.14.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
77
+ "model.layers.14.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
78
+ "model.layers.14.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
79
+ "model.layers.14.self_attn.k_proj",
80
+ "model.layers.14.self_attn.o_proj",
81
+ "model.layers.14.self_attn.q_proj",
82
+ "model.layers.14.self_attn.v_proj",
83
+ "model.layers.15.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
84
+ "model.layers.15.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
85
+ "model.layers.15.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
86
+ "model.layers.15.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
87
+ "model.layers.15.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
88
+ "model.layers.15.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
89
+ "model.layers.15.self_attn.k_proj",
90
+ "model.layers.15.self_attn.o_proj",
91
+ "model.layers.15.self_attn.q_proj",
92
+ "model.layers.15.self_attn.v_proj",
93
+ "model.layers.16.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
94
+ "model.layers.16.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
95
+ "model.layers.16.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
96
+ "model.layers.16.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
97
+ "model.layers.16.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
98
+ "model.layers.16.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
99
+ "model.layers.16.self_attn.k_proj",
100
+ "model.layers.16.self_attn.o_proj",
101
+ "model.layers.16.self_attn.q_proj",
102
+ "model.layers.16.self_attn.v_proj",
103
+ "model.layers.17.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
104
+ "model.layers.17.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
105
+ "model.layers.17.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
106
+ "model.layers.17.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
107
+ "model.layers.17.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
108
+ "model.layers.17.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
109
+ "model.layers.17.self_attn.k_proj",
110
+ "model.layers.17.self_attn.o_proj",
111
+ "model.layers.17.self_attn.q_proj",
112
+ "model.layers.17.self_attn.v_proj",
113
+ "model.layers.18.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
114
+ "model.layers.18.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
115
+ "model.layers.18.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
116
+ "model.layers.18.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
117
+ "model.layers.18.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
118
+ "model.layers.18.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
119
+ "model.layers.18.self_attn.k_proj",
120
+ "model.layers.18.self_attn.o_proj",
121
+ "model.layers.18.self_attn.q_proj",
122
+ "model.layers.18.self_attn.v_proj",
123
+ "model.layers.19.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
124
+ "model.layers.19.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
125
+ "model.layers.19.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
126
+ "model.layers.19.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
127
+ "model.layers.19.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
128
+ "model.layers.19.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
129
+ "model.layers.19.self_attn.k_proj",
130
+ "model.layers.19.self_attn.o_proj",
131
+ "model.layers.19.self_attn.q_proj",
132
+ "model.layers.19.self_attn.v_proj",
133
+ "model.layers.2.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
134
+ "model.layers.2.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
135
+ "model.layers.2.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
136
+ "model.layers.2.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
137
+ "model.layers.2.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
138
+ "model.layers.2.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
139
+ "model.layers.2.self_attn.k_proj",
140
+ "model.layers.2.self_attn.o_proj",
141
+ "model.layers.2.self_attn.q_proj",
142
+ "model.layers.2.self_attn.v_proj",
143
+ "model.layers.20.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
144
+ "model.layers.20.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
145
+ "model.layers.20.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
146
+ "model.layers.20.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
147
+ "model.layers.20.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
148
+ "model.layers.20.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
149
+ "model.layers.20.self_attn.k_proj",
150
+ "model.layers.20.self_attn.o_proj",
151
+ "model.layers.20.self_attn.q_proj",
152
+ "model.layers.20.self_attn.v_proj",
153
+ "model.layers.21.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
154
+ "model.layers.21.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
155
+ "model.layers.21.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
156
+ "model.layers.21.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
157
+ "model.layers.21.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
158
+ "model.layers.21.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
159
+ "model.layers.21.self_attn.k_proj",
160
+ "model.layers.21.self_attn.o_proj",
161
+ "model.layers.21.self_attn.q_proj",
162
+ "model.layers.21.self_attn.v_proj",
163
+ "model.layers.22.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
164
+ "model.layers.22.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
165
+ "model.layers.22.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
166
+ "model.layers.22.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
167
+ "model.layers.22.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
168
+ "model.layers.22.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
169
+ "model.layers.22.self_attn.k_proj",
170
+ "model.layers.22.self_attn.o_proj",
171
+ "model.layers.22.self_attn.q_proj",
172
+ "model.layers.22.self_attn.v_proj",
173
+ "model.layers.23.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
174
+ "model.layers.23.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
175
+ "model.layers.23.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
176
+ "model.layers.23.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
177
+ "model.layers.23.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
178
+ "model.layers.23.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
179
+ "model.layers.23.self_attn.k_proj",
180
+ "model.layers.23.self_attn.o_proj",
181
+ "model.layers.23.self_attn.q_proj",
182
+ "model.layers.23.self_attn.v_proj",
183
+ "model.layers.24.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
184
+ "model.layers.24.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
185
+ "model.layers.24.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
186
+ "model.layers.24.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
187
+ "model.layers.24.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
188
+ "model.layers.24.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
189
+ "model.layers.24.self_attn.k_proj",
190
+ "model.layers.24.self_attn.o_proj",
191
+ "model.layers.24.self_attn.q_proj",
192
+ "model.layers.24.self_attn.v_proj",
193
+ "model.layers.25.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
194
+ "model.layers.25.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
195
+ "model.layers.25.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
196
+ "model.layers.25.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
197
+ "model.layers.25.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
198
+ "model.layers.25.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
199
+ "model.layers.25.self_attn.k_proj",
200
+ "model.layers.25.self_attn.o_proj",
201
+ "model.layers.25.self_attn.q_proj",
202
+ "model.layers.25.self_attn.v_proj",
203
+ "model.layers.26.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
204
+ "model.layers.26.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
205
+ "model.layers.26.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
206
+ "model.layers.26.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
207
+ "model.layers.26.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
208
+ "model.layers.26.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
209
+ "model.layers.26.self_attn.k_proj",
210
+ "model.layers.26.self_attn.o_proj",
211
+ "model.layers.26.self_attn.q_proj",
212
+ "model.layers.26.self_attn.v_proj",
213
+ "model.layers.27.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
214
+ "model.layers.27.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
215
+ "model.layers.27.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
216
+ "model.layers.27.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
217
+ "model.layers.27.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
218
+ "model.layers.27.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
219
+ "model.layers.27.self_attn.k_proj",
220
+ "model.layers.27.self_attn.o_proj",
221
+ "model.layers.27.self_attn.q_proj",
222
+ "model.layers.27.self_attn.v_proj",
223
+ "model.layers.28.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
224
+ "model.layers.28.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
225
+ "model.layers.28.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
226
+ "model.layers.28.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
227
+ "model.layers.28.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
228
+ "model.layers.28.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
229
+ "model.layers.28.self_attn.k_proj",
230
+ "model.layers.28.self_attn.o_proj",
231
+ "model.layers.28.self_attn.q_proj",
232
+ "model.layers.28.self_attn.v_proj",
233
+ "model.layers.29.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
234
+ "model.layers.29.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
235
+ "model.layers.29.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
236
+ "model.layers.29.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
237
+ "model.layers.29.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
238
+ "model.layers.29.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
239
+ "model.layers.29.self_attn.k_proj",
240
+ "model.layers.29.self_attn.o_proj",
241
+ "model.layers.29.self_attn.q_proj",
242
+ "model.layers.29.self_attn.v_proj",
243
+ "model.layers.3.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
244
+ "model.layers.3.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
245
+ "model.layers.3.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
246
+ "model.layers.3.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
247
+ "model.layers.3.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
248
+ "model.layers.3.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
249
+ "model.layers.3.self_attn.k_proj",
250
+ "model.layers.3.self_attn.o_proj",
251
+ "model.layers.3.self_attn.q_proj",
252
+ "model.layers.3.self_attn.v_proj",
253
+ "model.layers.30.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
254
+ "model.layers.30.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
255
+ "model.layers.30.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
256
+ "model.layers.30.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
257
+ "model.layers.30.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
258
+ "model.layers.30.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
259
+ "model.layers.30.self_attn.k_proj",
260
+ "model.layers.30.self_attn.o_proj",
261
+ "model.layers.30.self_attn.q_proj",
262
+ "model.layers.30.self_attn.v_proj",
263
+ "model.layers.31.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
264
+ "model.layers.31.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
265
+ "model.layers.31.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
266
+ "model.layers.31.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
267
+ "model.layers.31.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
268
+ "model.layers.31.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
269
+ "model.layers.31.self_attn.k_proj",
270
+ "model.layers.31.self_attn.o_proj",
271
+ "model.layers.31.self_attn.q_proj",
272
+ "model.layers.31.self_attn.v_proj",
273
+ "model.layers.4.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
274
+ "model.layers.4.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
275
+ "model.layers.4.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
276
+ "model.layers.4.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
277
+ "model.layers.4.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
278
+ "model.layers.4.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
279
+ "model.layers.4.self_attn.k_proj",
280
+ "model.layers.4.self_attn.o_proj",
281
+ "model.layers.4.self_attn.q_proj",
282
+ "model.layers.4.self_attn.v_proj",
283
+ "model.layers.5.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
284
+ "model.layers.5.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
285
+ "model.layers.5.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
286
+ "model.layers.5.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
287
+ "model.layers.5.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
288
+ "model.layers.5.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
289
+ "model.layers.5.self_attn.k_proj",
290
+ "model.layers.5.self_attn.o_proj",
291
+ "model.layers.5.self_attn.q_proj",
292
+ "model.layers.5.self_attn.v_proj",
293
+ "model.layers.6.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
294
+ "model.layers.6.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
295
+ "model.layers.6.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
296
+ "model.layers.6.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
297
+ "model.layers.6.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
298
+ "model.layers.6.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
299
+ "model.layers.6.self_attn.k_proj",
300
+ "model.layers.6.self_attn.o_proj",
301
+ "model.layers.6.self_attn.q_proj",
302
+ "model.layers.6.self_attn.v_proj",
303
+ "model.layers.7.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
304
+ "model.layers.7.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
305
+ "model.layers.7.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
306
+ "model.layers.7.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
307
+ "model.layers.7.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
308
+ "model.layers.7.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
309
+ "model.layers.7.self_attn.k_proj",
310
+ "model.layers.7.self_attn.o_proj",
311
+ "model.layers.7.self_attn.q_proj",
312
+ "model.layers.7.self_attn.v_proj",
313
+ "model.layers.8.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
314
+ "model.layers.8.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
315
+ "model.layers.8.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
316
+ "model.layers.8.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
317
+ "model.layers.8.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
318
+ "model.layers.8.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
319
+ "model.layers.8.self_attn.k_proj",
320
+ "model.layers.8.self_attn.o_proj",
321
+ "model.layers.8.self_attn.q_proj",
322
+ "model.layers.8.self_attn.v_proj",
323
+ "model.layers.9.mlp.deepspeed_moe.experts.deepspeed_experts.0.down_proj",
324
+ "model.layers.9.mlp.deepspeed_moe.experts.deepspeed_experts.0.gate_proj",
325
+ "model.layers.9.mlp.deepspeed_moe.experts.deepspeed_experts.0.up_proj",
326
+ "model.layers.9.mlp.deepspeed_moe.experts.deepspeed_experts.1.down_proj",
327
+ "model.layers.9.mlp.deepspeed_moe.experts.deepspeed_experts.1.gate_proj",
328
+ "model.layers.9.mlp.deepspeed_moe.experts.deepspeed_experts.1.up_proj",
329
+ "model.layers.9.self_attn.k_proj",
330
+ "model.layers.9.self_attn.o_proj",
331
+ "model.layers.9.self_attn.q_proj",
332
+ "model.layers.9.self_attn.v_proj"
333
+ ],
334
+ "lora_alpha": 32,
335
+ "lora_dropout": 0.05,
336
+ "fan_in_fan_out": false,
337
+ "bias": "none",
338
+ "use_rslora": false,
339
+ "modules_to_save": null,
340
+ "init_lora_weights": true,
341
+ "layers_to_transform": null,
342
+ "layers_pattern": null,
343
+ "rank_pattern": {},
344
+ "alpha_pattern": {},
345
+ "megatron_config": null,
346
+ "megatron_core": "megatron.core",
347
+ "loftq_config": {},
348
+ "use_dora": false,
349
+ "layer_replication": null
350
+ },
351
+ "epoch": 11,
352
+ "stage": "B",
353
+ "seed": 42,
354
+ "opt_steps": 937
355
+ }
assets/fig_architecture.png ADDED

Git LFS Details

  • SHA256: 6569b4881c1685b8eaaed1c71038d237a41664e2462f6b43f155620e0a520fae
  • Pointer size: 131 Bytes
  • Size of remote file: 107 kB
assets/fig_results.png ADDED
assets/fig_stages.png ADDED
trainable.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ea13d55d06167d32b4e08abf26a1d22c7d82c259a83129b6c2b353f390a49c6b
3
+ size 573948474