KenteaKe11 commited on
Commit
1e845fd
·
verified ·
1 Parent(s): 24c51a5

§3.1 — 3.1_c_vs_baseline_v4

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. 3.1_c_vs_baseline_v4/REPORT.md +326 -0
  2. 3.1_c_vs_baseline_v4/arms/ac_e4/REPORT.md +44 -0
  3. 3.1_c_vs_baseline_v4/arms/ac_e4/e_clinical_corr.csv +5 -0
  4. 3.1_c_vs_baseline_v4/arms/ac_e4/e_e_corr.csv +5 -0
  5. 3.1_c_vs_baseline_v4/arms/ac_e4/e_heatmap.png +3 -0
  6. 3.1_c_vs_baseline_v4/arms/ac_e4/route_verdict.json +24 -0
  7. 3.1_c_vs_baseline_v4/arms/ac_e4/summary.json +202 -0
  8. 3.1_c_vs_baseline_v4/arms/baseline_infer/REPORT.md +15 -0
  9. 3.1_c_vs_baseline_v4/arms/baseline_infer/summary.json +78 -0
  10. 3.1_c_vs_baseline_v4/arms/c1_e4/REPORT.md +44 -0
  11. 3.1_c_vs_baseline_v4/arms/c1_e4/e_clinical_corr.csv +5 -0
  12. 3.1_c_vs_baseline_v4/arms/c1_e4/e_e_corr.csv +5 -0
  13. 3.1_c_vs_baseline_v4/arms/c1_e4/e_heatmap.png +3 -0
  14. 3.1_c_vs_baseline_v4/arms/c1_e4/route_verdict.json +24 -0
  15. 3.1_c_vs_baseline_v4/arms/c1_e4/summary.json +202 -0
  16. 3.1_c_vs_baseline_v4/arms/c2_e4/REPORT.md +44 -0
  17. 3.1_c_vs_baseline_v4/arms/c2_e4/e_clinical_corr.csv +5 -0
  18. 3.1_c_vs_baseline_v4/arms/c2_e4/e_e_corr.csv +5 -0
  19. 3.1_c_vs_baseline_v4/arms/c2_e4/e_heatmap.png +3 -0
  20. 3.1_c_vs_baseline_v4/arms/c2_e4/route_verdict.json +24 -0
  21. 3.1_c_vs_baseline_v4/arms/c2_e4/summary.json +202 -0
  22. 3.1_c_vs_baseline_v4/best.json +40 -0
  23. 3.1_c_vs_baseline_v4/config.resolved.yaml +252 -0
  24. 3.1_c_vs_baseline_v4/logs/section3_20261105_204504.log +539 -0
  25. 3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_e_clinical_corr.csv +5 -0
  26. 3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_e_e_corr.csv +5 -0
  27. 3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_e_heatmap.png +3 -0
  28. 3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_expert_labels_lift.csv +5 -0
  29. 3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_expert_labels_mean.csv +5 -0
  30. 3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_route_verdict.json +24 -0
  31. 3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_sim_param_cos.csv +5 -0
  32. 3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_sim_route_corr.csv +5 -0
  33. 3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_sim_route_cos.csv +5 -0
  34. 3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_e_clinical_corr.csv +5 -0
  35. 3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_e_e_corr.csv +5 -0
  36. 3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_e_heatmap.png +3 -0
  37. 3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_expert_labels_lift.csv +5 -0
  38. 3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_expert_labels_mean.csv +5 -0
  39. 3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_route_verdict.json +24 -0
  40. 3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_sim_param_cos.csv +5 -0
  41. 3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_sim_route_corr.csv +5 -0
  42. 3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_sim_route_cos.csv +5 -0
  43. 3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_e_clinical_corr.csv +5 -0
  44. 3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_e_e_corr.csv +5 -0
  45. 3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_e_heatmap.png +3 -0
  46. 3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_expert_labels_lift.csv +5 -0
  47. 3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_expert_labels_mean.csv +5 -0
  48. 3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_route_verdict.json +24 -0
  49. 3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_sim_param_cos.csv +5 -0
  50. 3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_sim_route_corr.csv +5 -0
3.1_c_vs_baseline_v4/REPORT.md ADDED
@@ -0,0 +1,326 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # §3.1 — Cls head · EC-Router · MoE-DoRA
2
+
3
+ **Câu hỏi.** Ở E cố định, giá trị capacity c nào tốt nhất — (c=1), (c=2) hay (c=a: Adaptive Capacity) — và tất cả so với Baseline (Gemma 3 gốc) thế nào, xét theo text metric, clinical metric và E-Clinical (Pearson S, S·Sᵀ)? Các expert có chuyên biệt hay sụp đổ về một DoRA?
4
+
5
+ > Greedy · mọi PEFT = QDoRA r=4 · loss = CE text + γ·CE cls (γ = 0.1, huấn luyện cùng lúc) · nhãn cls sinh lazy từ CheXbert · 6400 ảnh test. Các bảng thiết lập, điểm, Cls head và Route lấy trực tiếp từ báo cáo con của từng arm (`arms/<arm>/REPORT.md`).
6
+
7
+ ## Kết luận nhanh
8
+
9
+ - **Xếp hạng #1 theo hạng trung bình (gợi ý, chưa qua kiểm định):** arm `ac_e4` — hạng lâm sàng TB 1.00, hạng văn bản TB 1.12 → c = a, E = 4 (`best.json`).
10
+ - **Chuyên biệt hay sụp đổ:** không arm nào vượt ngưỡng sụp đổ. Chi tiết từng arm ở mục E-Clinical.
11
+ - Một khác biệt chỉ có nghĩa khi p_perm < α và CI bootstrap loại 0 (xem mục Friedman và hậu kiểm).
12
+
13
+ ## Báo cáo con
14
+
15
+ - [MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉)](arms/c1_e4/REPORT.md)
16
+ - [MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉)](arms/c2_e4/REPORT.md)
17
+ - [MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉)](arms/ac_e4/REPORT.md)
18
+ - [Baseline · Gemma 3 gốc](arms/baseline_infer/REPORT.md)
19
+
20
+ ## Thiết lập các arm
21
+
22
+ | Arm | Decoder | E | c | k mỗi expert | Cls head | Tham số học | Phút |
23
+ |---|---|---|---|---|---|---|---|
24
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) | moe | 4 | 1 | ⌈c·N/E⌉ | đồng huấn luyện | 35,381,602 | 608.6 |
25
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | moe | 4 | 2 | ⌈c·N/E⌉ | đồng huấn luyện | 35,381,602 | 607.5 |
26
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | moe | 4 | adaptive | ⌈max(N/E, e^H_i)⌉ | đồng huấn luyện | 35,381,602 | 610.1 |
27
+ | Baseline · Gemma 3 gốc | — | — | — | — | — | 0 | — |
28
+
29
+ ## Điểm
30
+
31
+ ### Clinical metric
32
+
33
+ | Mô hình | CX-F1 | RG-E | RG-P | RG-C |
34
+ | :--- | ---: | ---: | ---: | ---: |
35
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) | 0.2254<br><sub>±0.454</sub> | 0.3413<br><sub>±0.191</sub> | 0.3237<br><sub>±0.189</sub> | 0.2659<br><sub>±0.179</sub> |
36
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | 0.2306<br><sub>±0.453</sub> | 0.3420<br><sub>±0.191</sub> | 0.3242<br><sub>±0.189</sub> | 0.2660<br><sub>±0.179</sub> |
37
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | **0.2342**<br><sub>±0.453</sub> | **0.3437**<br><sub>±0.190</sub> | **0.3257**<br><sub>±0.188</sub> | **0.2674**<br><sub>±0.179</sub> |
38
+ | Baseline · Gemma 3 gốc | 0.2144<br><sub>±0.257</sub> | 0.2212<br><sub>±0.110</sub> | 0.2027<br><sub>±0.104</sub> | 0.1199<br><sub>±0.084</sub> |
39
+
40
+ ### Text metric
41
+
42
+ | Mô hình | B-1 | B-2 | B-3 | B-4 | R-1 | R-2 | R-L | BS-F1 |
43
+ | :--- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
44
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) | 0.3035<br><sub>±0.215</sub> | 0.2323<br><sub>±0.187</sub> | 0.1799<br><sub>±0.169</sub> | 0.1444<br><sub>±0.158</sub> | 0.4312<br><sub>±0.175</sub> | 0.1986<br><sub>±0.164</sub> | 0.3467<br><sub>±0.168</sub> | 0.8946<br><sub>±0.031</sub> |
45
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | 0.3050<br><sub>±0.214</sub> | 0.2334<br><sub>±0.186</sub> | 0.1807<br><sub>±0.169</sub> | 0.1450<br><sub>±0.157</sub> | 0.4323<br><sub>±0.174</sub> | 0.1994<br><sub>±0.164</sub> | 0.3471<br><sub>±0.167</sub> | 0.8944<br><sub>±0.031</sub> |
46
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | 0.3074<br><sub>±0.213</sub> | **0.2347**<br><sub>±0.186</sub> | **0.1815**<br><sub>±0.168</sub> | **0.1456**<br><sub>±0.157</sub> | **0.4341**<br><sub>±0.172</sub> | **0.2000**<br><sub>±0.163</sub> | **0.3481**<br><sub>±0.167</sub> | **0.8946**<br><sub>±0.031</sub> |
47
+ | Baseline · Gemma 3 gốc | **0.3233**<br><sub>±0.101</sub> | 0.1958<br><sub>±0.074</sub> | 0.1177<br><sub>±0.059</sub> | 0.0691<br><sub>±0.044</sub> | 0.3291<br><sub>±0.088</sub> | 0.0887<br><sub>±0.054</sub> | 0.2104<br><sub>±0.066</sub> | 0.8640<br><sub>±0.020</sub> |
48
+
49
+ ## Cls head (nhãn CheXbert dương = positive|uncertain)
50
+
51
+ | Arm | Precision | Recall | micro-F1 | n |
52
+ |---|---|---|---|---|
53
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) | 0.6009 | 0.0113 | 0.0223 | 6400 |
54
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | 0.5633 | 0.0151 | 0.0294 | 6400 |
55
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | 0.5556 | 0.0144 | 0.0281 | 6400 |
56
+
57
+ ## Xếp hạng cấu hình
58
+
59
+ Mọi metric lâm sàng (CheXbert F1, RadGraph) đều là metric chính: xếp theo **hạng lâm sàng trung bình** (nhỏ hơn = tốt hơn); khi hạng lâm sàng ngang nhau thì xét **hạng metric sinh văn bản** (BLEU, ROUGE, BERTScore).
60
+
61
+ | Xếp hạng | Arm | Hạng lâm sàng TB | Hạng văn bản TB | Hoà lâm sàng |
62
+ |---|---|---|---|---|
63
+ | 1 | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | 1.00 | 1.12 | — |
64
+ | 2 | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | 2.00 | 2.25 | — |
65
+ | 3 | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) | 3.00 | 3.00 | — |
66
+ | 4 | Baseline · Gemma 3 gốc | 4.00 | 3.62 | — |
67
+
68
+ ## Kiểm định Friedman + hậu kiểm
69
+
70
+ H0: **4 arm có cùng phân phối hạng trong mỗi khối** (khối = 6400 ảnh test). Friedman chạy cho **mọi metric**; hậu kiểm Wilcoxon + Holm chỉ chạy trên metric đã bác H0 (kết luận cần |rank-biserial| ≥ 0.1). Mọi bảng chi tiết nằm trong các khối **▸ bấm để mở/thu gọn** bên dưới.
71
+
72
+ ### Bảng tóm tắt
73
+
74
+ | Metric | Friedman | Kendall W | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | Baseline · Gemma 3 gốc | Kết luận |
75
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
76
+ | `bleu1` | ✓ bác H0 (W thấp) | 0.028 | 2.61 ▲0 ▼2 | 2.60 ▲0 ▼2 | 2.55 ▲2 ▼1 | **2.24** ▲3 ▼0 | Baseline · Gemma 3 gốc — thắng 3/3 arm |
77
+ | `bleu2` | ✓ bác H0 (W thấp) | 0.003 | 2.56 ▲1 ▼1 | 2.53 ▲1 ▼1 | 2.48 ▲3 ▼0 | **2.43** ▲0 ▼3 | Baseline · Gemma 3 gốc hạng TB tốt nhất, chưa thắng rõ arm nào |
78
+ | `bleu3` | ✓ bác H0 (W thấp) | 0.004 | 2.50 ▲1 ▼2 | 2.47 ▲2 ▼1 | **2.43** ▲3 ▼0 | 2.59 ▲0 ▼3 | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) — thắng 3/3 arm |
79
+ | `bleu4` | ✓ bác H0 (W thấp) | 0.030 | 2.45 ▲1 ▼2 | 2.41 ▲2 ▼1 | **2.37** ▲3 ▼0 | 2.77 ▲0 ▼3 | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) — thắng 3/3 arm |
80
+ | `rouge1` | ✓ bác H0 | 0.097 | 2.37 ▲1 ▼1 | 2.34 ▲1 ▼1 | **2.30** ▲3 ▼0 | 2.99 ▲0 ▼3 | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) — thắng 3/3 arm |
81
+ | `rouge2` | ✓ bác H0 | 0.239 | 2.27 ▲1 ▼2 | 2.23 ▲2 ▼0 | **2.23** ▲2 ▼0 | 3.27 ▲0 ▼3 | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) — thắng 2/3 arm |
82
+ | `rougeL` | ✓ bác H0 | 0.304 | 2.23 ▲1 ▼1 | 2.21 ▲1 ▼0 | **2.19** ▲2 ▼0 | 3.38 ▲0 ▼3 | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) — thắng 2/3 arm |
83
+ | `bertscore_f1` | ✓ bác H0 | 0.597 | **2.07** ▲2 ▼1 | 2.10 ▲1 ▼1 | 2.10 ▲2 ▼0 | 3.73 ▲0 ▼3 | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) — thắng 2/3 arm |
84
+ | `chexbert_f1` | ✓ bác H0 (W thấp) | 0.002 | 2.49 ▲1 ▼2 | 2.48 ▲2 ▼0 | **2.47** ▲2 ▼0 | 2.55 ▲0 ▼3 | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) — thắng 2/3 arm |
85
+ | `radgraph_f1_entity` | ✓ bác H0 | 0.178 | 2.30 ▲1 ▼1 | 2.27 ▲1 ▼0 | **2.26** ▲2 ▼0 | 3.17 ▲0 ▼3 | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) — thắng 2/3 arm |
86
+ | `radgraph_f1_partial` | ✓ bác H0 | 0.187 | 2.29 ▲1 ▼0 | 2.27 ▲1 ▼0 | **2.26** ▲1 ▼0 | 3.18 ▲0 ▼3 | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) — thắng 1/3 arm |
87
+ | `radgraph_f1_complete` | ✓ bác H0 | 0.342 | 2.20 ▲1 ▼0 | **2.19** ▲1 ▼0 | 2.19 ▲1 ▼0 | 3.42 ▲0 ▼3 | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) — thắng 1/3 arm |
88
+
89
+ *Ô arm = hạng trung bình (1 = tốt nhất, đậm = tốt nhất). ▲w ▼l = số arm mà arm này thắng / số arm thắng nó, tính ở hậu kiểm (chỉ hiện ở metric đã bác H0). Hạng TB là mô tả; kết luận hơn–thua chỉ đến từ ▲▼.*
90
+
91
+ <details>
92
+ <summary>Friedman chi tiết (χ², F Iman-Davenport, p, p Holm, Kendall W)</summary>
93
+
94
+ | Metric | χ² | F (Iman-Davenport) | p | p Holm | Kendall W | khối trùng hoàn toàn | bác H0? |
95
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
96
+ | `bleu1` | 536.3 | 183.9 (3,19197) | <0.001 | <0.001 | 0.0279 | 0 (0%) | ✓ |
97
+ | `bleu2` | 56.5 | 18.9 (3,19197) | <0.001 | <0.001 | 0.0029 | 0 (0%) | ✓ |
98
+ | `bleu3` | 79.5 | 26.6 (3,19197) | <0.001 | <0.001 | 0.0041 | 0 (0%) | ✓ |
99
+ | `bleu4` | 573.6 | 197.0 (3,19197) | <0.001 | <0.001 | 0.0299 | 0 (0%) | ✓ |
100
+ | `rouge1` | 1855.0 | 684.4 (3,19197) | <0.001 | <0.001 | 0.0966 | 1 (0%) | ✓ |
101
+ | `rouge2` | 4582.3 | 2005.9 (3,19197) | <0.001 | <0.001 | 0.2387 | 15 (0%) | ✓ |
102
+ | `rougeL` | 5838.9 | 2796.4 (3,19197) | <0.001 | <0.001 | 0.3041 | 1 (0%) | ✓ |
103
+ | `bertscore_f1` | 11456.2 | 9466.7 (3,19197) | <0.001 | <0.001 | 0.5967 | 0 (0%) | ✓ |
104
+ | `chexbert_f1` | 45.6 | 15.2 (3,19197) | <0.001 | <0.001 | 0.0024 | 2767 (43%) | ✓ |
105
+ | `radgraph_f1_entity` | 3414.7 | 1384.3 (3,19197) | <0.001 | <0.001 | 0.1779 | 33 (1%) | ✓ |
106
+ | `radgraph_f1_partial` | 3582.2 | 1467.7 (3,19197) | <0.001 | <0.001 | 0.1866 | 37 (1%) | ✓ |
107
+ | `radgraph_f1_complete` | 6566.3 | 3325.8 (3,19197) | <0.001 | <0.001 | 0.3420 | 81 (1%) | ✓ |
108
+
109
+ Kendall W là effect size của kiểm định tổng thể (0 = các ảnh xếp hạng không nhất quán, 1 = mọi ảnh xếp hạng y hệt nhau); với n = 6400, p nhỏ không nói lên nhiều — đọc W trước. *Khối trùng hoàn toàn* = ảnh mà mọi arm được đúng cùng một điểm (hay gặp ở BLEU-4).
110
+
111
+ > ⚠ `bleu1`: Kendall W = 0.0279 < 0.05: mức đồng thuận về thứ hạng giữa 6400 khối là RẤT THẤP, dù p có nhỏ tới đâu. Với n lớn thế này, p-value không còn là tiêu chí quyết định — đọc W, hạng trung bình và rank-biserial của hậu kiểm trước khi kết luận biến thể nào 'hơn'.
112
+
113
+ > ⚠ `bleu2`: Kendall W = 0.0029 < 0.05: mức đồng thuận về thứ hạng giữa 6400 khối là RẤT THẤP, dù p có nhỏ tới đâu. Với n lớn thế này, p-value không còn là tiêu chí quyết định — đọc W, hạng trung bình và rank-biserial của hậu kiểm trước khi kết luận biến thể nào 'hơn'.
114
+
115
+ > ⚠ `bleu3`: Kendall W = 0.0041 < 0.05: mức đồng thuận về thứ hạng giữa 6400 khối là RẤT THẤP, dù p có nhỏ tới đâu. Với n lớn thế này, p-value không còn là tiêu chí quyết định — đọc W, hạng trung bình và rank-biserial của hậu kiểm trước khi kết luận biến thể nào 'hơn'.
116
+
117
+ > ⚠ `bleu4`: Kendall W = 0.0299 < 0.05: mức đồng thuận về thứ hạng giữa 6400 khối là RẤT THẤP, dù p có nhỏ tới đâu. Với n lớn thế này, p-value không còn là tiêu chí quyết định — đọc W, hạng trung bình và rank-biserial của hậu kiểm trước khi kết luận biến thể nào 'hơn'.
118
+
119
+ > ⚠ `chexbert_f1`: Kendall W = 0.0024 < 0.05: mức đồng thuận về thứ hạng giữa 6400 khối là RẤT THẤP, dù p có nhỏ tới đâu. Với n lớn thế này, p-value không còn là tiêu chí quyết định — đọc W, hạng trung bình và rank-biserial của hậu kiểm trước khi kết luận biến thể nào 'hơn'.
120
+
121
+ </details>
122
+
123
+ Hậu kiểm: 72 kiểm định = 6 cặp × 12 metric đã bác H0, Wilcoxon ghép cặp (`zero_method=pratt`), Holm toàn cục; hiệu số **Δ = A − B**. **59** đạt ý nghĩa thống kê, **56** trong số đó đáng kể.
124
+
125
+ <details>
126
+ <summary>Hậu kiểm `chexbert_f1` — 6 cặp, 5 khác biệt đáng kể</summary>
127
+
128
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
129
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
130
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.1611 | 1932/2806/1662 | 14.1 | <0.001 | <0.001 | 0.241 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
131
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1595 | 1918/2798/1684 | 13.8 | <0.001 | <0.001 | 0.234 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
132
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1567 | 1887/2809/1704 | 13.2 | <0.001 | <0.001 | 0.225 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
133
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0044 | 38/6247/115 | -6.2 | <0.001 | <0.001 | -0.503 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
134
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0027 | 45/6269/86 | -3.6 | <0.001 | 0.006 | -0.313 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
135
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0016 | 66/6242/92 | -2.1 | 0.038 | 0.38 | -0.165 | — |
136
+
137
+ </details>
138
+
139
+ <details>
140
+ <summary>Hậu kiểm `bleu1` — 6 cặp, 5 khác biệt đáng kể</summary>
141
+
142
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
143
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
144
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | -0.0192 | 2617/0/3783 | -9.9 | <0.001 | <0.001 | -0.143 | ✓ Baseline · Gemma 3 gốc hơn |
145
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | -0.0174 | 2644/0/3756 | -9.4 | <0.001 | <0.001 | -0.136 | ✓ Baseline · Gemma 3 gốc hơn |
146
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | -0.0146 | 2662/0/3738 | -8.7 | <0.001 | <0.001 | -0.125 | ✓ Baseline · Gemma 3 gốc hơn |
147
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0046 | 436/5300/664 | -7.3 | <0.001 | <0.001 | -0.221 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
148
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0028 | 401/5380/619 | -6.9 | <0.001 | <0.001 | -0.218 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
149
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0018 | 511/5356/533 | -0.9 | 0.34 | 1.00 | -0.029 | — |
150
+
151
+ </details>
152
+
153
+ <details>
154
+ <summary>Hậu kiểm `bleu2` — 6 cặp, 5 khác biệt đáng kể</summary>
155
+
156
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
157
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
158
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.0400 | 3071/0/3329 | 8.6 | <0.001 | <0.001 | 0.124 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
159
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0028 | 424/5292/684 | -8.1 | <0.001 | <0.001 | -0.244 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
160
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.0385 | 3048/0/3352 | 8.0 | <0.001 | <0.001 | 0.115 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
161
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.0373 | 3012/0/3388 | 7.4 | <0.001 | <0.001 | 0.107 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
162
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0016 | 416/5367/617 | -6.3 | <0.001 | <0.001 | -0.196 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
163
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0012 | 481/5350/569 | -2.9 | 0.004 | 0.054 | -0.088 | — |
164
+
165
+ </details>
166
+
167
+ <details>
168
+ <summary>Hậu kiểm `bleu3` — 6 cặp, 6 khác biệt đáng kể</summary>
169
+
170
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
171
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
172
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.0647 | 3421/0/2979 | 21.7 | <0.001 | <0.001 | 0.313 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
173
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.0638 | 3399/0/3001 | 21.1 | <0.001 | <0.001 | 0.304 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
174
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.0629 | 3374/0/3026 | 20.5 | <0.001 | <0.001 | 0.296 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
175
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0019 | 426/5288/686 | -8.0 | <0.001 | <0.001 | -0.239 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
176
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0010 | 446/5362/592 | -4.6 | <0.001 | <0.001 | -0.142 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
177
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0009 | 461/5346/593 | -4.1 | <0.001 | <0.001 | -0.128 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
178
+
179
+ </details>
180
+
181
+ <details>
182
+ <summary>Hậu kiểm `bleu4` — 6 cặp, 6 khác biệt đáng kể</summary>
183
+
184
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
185
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
186
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.0771 | 3803/0/2597 | 32.7 | <0.001 | <0.001 | 0.472 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
187
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.0764 | 3783/0/2617 | 32.1 | <0.001 | <0.001 | 0.464 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
188
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.0758 | 3746/0/2654 | 31.5 | <0.001 | <0.001 | 0.455 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
189
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0013 | 421/5287/692 | -8.2 | <0.001 | <0.001 | -0.246 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
190
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0006 | 452/5346/602 | -4.6 | <0.001 | <0.001 | -0.143 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
191
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0007 | 448/5361/591 | -4.4 | <0.001 | <0.001 | -0.138 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
192
+
193
+ </details>
194
+
195
+ <details>
196
+ <summary>Hậu kiểm `rouge1` — 6 cặp, 5 khác biệt đáng kể</summary>
197
+
198
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
199
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
200
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.1059 | 4285/4/2111 | 38.3 | <0.001 | <0.001 | 0.552 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
201
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1039 | 4243/2/2155 | 37.3 | <0.001 | <0.001 | 0.538 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
202
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1028 | 4214/2/2184 | 36.7 | <0.001 | <0.001 | 0.530 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
203
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0031 | 444/5297/659 | -6.7 | <0.001 | <0.001 | -0.203 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
204
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0020 | 440/5374/586 | -4.7 | <0.001 | <0.001 | -0.147 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
205
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0011 | 470/5361/569 | -3.1 | 0.002 | 0.029 | -0.097 | ≈ có ý nghĩa nhưng KHÔNG đáng kể |
206
+
207
+ </details>
208
+
209
+ <details>
210
+ <summary>Hậu kiểm `rouge2` — 6 cặp, 5 khác biệt đáng kể</summary>
211
+
212
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
213
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
214
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1105 | 4810/21/1569 | 51.2 | <0.001 | <0.001 | 0.740 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
215
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1111 | 4841/19/1540 | 51.8 | <0.001 | <0.001 | 0.747 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
216
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.1118 | 4865/20/1515 | 52.3 | <0.001 | <0.001 | 0.754 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
217
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0013 | 475/5325/600 | -3.9 | <0.001 | 0.002 | -0.119 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
218
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0007 | 445/5386/569 | -3.8 | <0.001 | 0.003 | -0.120 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
219
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0006 | 495/5398/507 | -0.5 | 0.62 | 1.00 | -0.016 | — |
220
+
221
+ </details>
222
+
223
+ <details>
224
+ <summary>Hậu kiểm `rougeL` — 6 cặp, 4 khác biệt đáng kể</summary>
225
+
226
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
227
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
228
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1366 | 5036/2/1362 | 55.8 | <0.001 | <0.001 | 0.806 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
229
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1370 | 5068/2/1330 | 56.2 | <0.001 | <0.001 | 0.812 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
230
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.1380 | 5093/2/1305 | 56.8 | <0.001 | <0.001 | 0.820 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
231
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0014 | 483/5292/625 | -4.4 | <0.001 | <0.001 | -0.131 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
232
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0010 | 480/5363/557 | -2.5 | 0.013 | 0.15 | -0.077 | — |
233
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0004 | 501/5347/552 | -1.6 | 0.10 | 0.73 | -0.050 | — |
234
+
235
+ </details>
236
+
237
+ <details>
238
+ <summary>Hậu kiểm `bertscore_f1` — 6 cặp, 5 khác biệt đáng kể</summary>
239
+
240
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
241
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
242
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.0307 | 5837/0/563 | 65.6 | <0.001 | <0.001 | 0.947 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
243
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.0305 | 5802/0/598 | 65.2 | <0.001 | <0.001 | 0.941 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
244
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.0307 | 5828/0/572 | 65.6 | <0.001 | <0.001 | 0.947 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
245
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0000 | 661/5225/514 | 4.2 | <0.001 | <0.001 | 0.122 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
246
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | 0.0002 | 609/5313/478 | 4.0 | <0.001 | 0.001 | 0.121 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
247
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0002 | 542/5306/552 | -0.4 | 0.68 | 1.00 | -0.013 | — |
248
+
249
+ </details>
250
+
251
+ <details>
252
+ <summary>Hậu kiểm `radgraph_f1_entity` — 6 cặp, 4 khác biệt đáng kể</summary>
253
+
254
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
255
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
256
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1210 | 4582/38/1780 | 43.9 | <0.001 | <0.001 | 0.634 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
257
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1217 | 4594/40/1766 | 44.4 | <0.001 | <0.001 | 0.640 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
258
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.1234 | 4628/39/1733 | 45.1 | <0.001 | <0.001 | 0.652 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
259
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0024 | 471/5344/585 | -3.7 | <0.001 | 0.004 | -0.113 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
260
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0007 | 461/5381/558 | -3.0 | 0.003 | 0.044 | -0.093 | ≈ có ý nghĩa nhưng KHÔNG đáng kể |
261
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0017 | 476/5410/514 | -1.4 | 0.16 | 0.95 | -0.045 | — |
262
+
263
+ </details>
264
+
265
+ <details>
266
+ <summary>Hậu kiểm `radgraph_f1_partial` — 6 cặp, 3 khác biệt đáng kể</summary>
267
+
268
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
269
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
270
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1218 | 4610/44/1746 | 45.2 | <0.001 | <0.001 | 0.652 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
271
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1223 | 4637/48/1715 | 45.5 | <0.001 | <0.001 | 0.656 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
272
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.1239 | 4646/47/1707 | 46.2 | <0.001 | <0.001 | 0.666 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
273
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0021 | 478/5353/569 | -3.0 | 0.003 | 0.044 | -0.092 | ≈ có ý nghĩa nhưng KHÔNG đáng kể |
274
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0005 | 460/5393/547 | -2.6 | 0.008 | 0.098 | -0.083 | — |
275
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0016 | 481/5417/502 | -0.9 | 0.38 | 1.00 | -0.028 | — |
276
+
277
+ </details>
278
+
279
+ <details>
280
+ <summary>Hậu kiểm `radgraph_f1_complete` — 6 cặp, 3 khác biệt đáng kể</summary>
281
+
282
+ | Cặp (A vs B) | Δ = A − B | A thắng/hoà/thua | z | p Wilcoxon | p Holm | r_rb | kết luận |
283
+ | :--- | :--- | :--- | :--- | :--- | :--- | :--- | :--- |
284
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1472 | 5104/94/1202 | 55.4 | <0.001 | <0.001 | 0.799 | ✓ MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) hơn |
285
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs Baseline · Gemma 3 gốc | 0.1472 | 5111/103/1186 | 55.4 | <0.001 | <0.001 | 0.800 | ✓ MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) hơn |
286
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) vs Baseline · Gemma 3 gốc | 0.1487 | 5122/102/1176 | 56.0 | <0.001 | <0.001 | 0.808 | ✓ MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) hơn |
287
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | -0.0001 | 468/5397/535 | -2.0 | 0.046 | 0.42 | -0.063 | — |
288
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0015 | 497/5352/551 | -1.7 | 0.080 | 0.64 | -0.054 | — |
289
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) vs MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | -0.0014 | 495/5422/483 | 0.2 | 0.86 | 1.00 | 0.006 | — |
290
+
291
+ </details>
292
+
293
+ Toàn bộ số liệu: `metrics/friedman_tests.csv`, `metrics/posthoc_tests.csv`.
294
+
295
+
296
+ ## E-Clinical: Pearson (S) và S·Sᵀ — các Expert có chuyên biệt không?
297
+
298
+ Forward tập test (eval), route w ∈ R^E mỗi ảnh = tỉ lệ token prefix mà từng expert phụ trách (trung bình trọng số theo token). **S** (E×14) = Pearson giữa trọng số route của expert i qua các ảnh và nhãn dương CheXbert j; **S·Sᵀ** (E×E) = hai expert giống nhau khi cùng tương quan với cùng bệnh. *Sụp đổ* = cosine ΔW giữa các expert ≥ ngưỡng, HOẶC route tĩnh (trọng số gần như không đổi theo ảnh).
299
+
300
+ ### MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉)
301
+
302
+ **Kết luận:** có phân hoá giữa các expert — cosine tham số ΔW (TB ngoài đường chéo) **0.073** (ngưỡng 0.9); hệ số biến thiên của route qua ảnh **0.124** (tĩnh nếu < 0.05); 6400 ảnh test.
303
+
304
+ ![E-Clinical | E-E — MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉)](arms/c1_e4/e_heatmap.png)
305
+
306
+ ### MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉)
307
+
308
+ **Kết luận:** có phân hoá giữa các expert — cosine tham số ΔW (TB ngoài đường chéo) **0.108** (ngưỡng 0.9); hệ số biến thiên của route qua ảnh **0.099** (tĩnh nếu < 0.05); 6400 ảnh test.
309
+
310
+ ![E-Clinical | E-E — MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉)](arms/c2_e4/e_heatmap.png)
311
+
312
+ ### MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉)
313
+
314
+ **Kết luận:** có phân hoá giữa các expert — cosine tham số ΔW (TB ngoài đường chéo) **0.109** (ngưỡng 0.9); hệ số biến thiên của route qua ảnh **0.073** (tĩnh nếu < 0.05); 6400 ảnh test.
315
+
316
+ ![E-Clinical | E-E — MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉)](arms/ac_e4/e_heatmap.png)
317
+
318
+ ## Giả định cần biết
319
+
320
+ - `best.json` chỉ là gợi ý theo hạng lâm sàng TB (hoà ⇒ hạng văn bản TB), không qua kiểm định; quyết định chọn cấu hình là của người dùng.
321
+ - EC-Router nhân quả (Expert-Choice): RMSNorm → Linear D→E; tại vị trí t mỗi expert chọn token nếu điểm thuộc top-k_t của các token ≤ t, k_t = ⌈c·n_t/E⌉; w theo từng token, chuẩn hoá giữa các expert đã chọn; token không expert nào chọn dùng trọng số nền đóng băng. Router và Cls head nhận toàn bộ token (ảnh + prompt + câu trả lời) nhưng nhân quả như decoder. Không nhiễu, không aux loss.
322
+ - Cls head nhân quả đọc hidden cuối của decoder ở mọi vị trí (suy luận: vị trí cuối báo cáo), huấn luyện cùng lúc (không pretrain, không đóng băng).
323
+ - Baseline là Gemma 3 gốc (không huấn luyện, greedy, cùng prompt/ảnh/tập test).
324
+ - c = a (Adaptive Capacity): k_i = ⌈max(N/E, e^{H_i})⌉ với H_i là entropy của softmax nhân quả theo token của expert i; E là siêu tham số duy nhất.
325
+ - 3.2 chọn c từ điểm của 3.1 trên CÙNG tập test (hạng lâm sàng TB, hoà ⇒ hạng văn bản), rồi dò E: lựa chọn c đó là một nguồn lạc quan nhẹ cho các arm của 3.2. Arm E của 3.1 và Baseline được dùng lại nguyên (cùng predictions, cùng điểm).
326
+ - Data (exp 3.0): train 12800 · val 128 · test 6400, 1 bệnh nhân – 1 mẫu, nhãn dương cân bằng giữa ba tập (xem `3.0_data_v4/DATA_REPORT.md`). Train 400 bước × batch 32, val mỗi 40 bước, patience 3, min_delta (xem config); 4 tập bootstrap × 6400 mẫu.
3.1_c_vs_baseline_v4/arms/ac_e4/REPORT.md ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # §3.1 — MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉)
2
+
3
+ Greedy · QDoRA r=4 · 6400 ảnh test. Báo cáo con của riêng arm này; báo cáo cuối của mục chỉ đọc lại `summary.json` cùng thư mục.
4
+
5
+ ## Thiết lập
6
+
7
+ | Arm | Decoder | E | c | k mỗi expert | Cls head | Tham số học | Phút |
8
+ |---|---|---|---|---|---|---|---|
9
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | moe | 4 | adaptive | ⌈max(N/E, e^H_i)⌉ | đồng huấn luyện | 35,381,602 | 610.1 |
10
+
11
+ ## Điểm
12
+
13
+ | Mô hình | B-1 | B-2 | B-3 | B-4 | R-1 | R-2 | R-L | BS-F1 | CX-F1 | RG-E | RG-P | RG-C |
14
+ | :--- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
15
+ | MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉) | **0.3074**<br><sub>±0.213</sub> | **0.2347**<br><sub>±0.186</sub> | **0.1815**<br><sub>±0.168</sub> | **0.1456**<br><sub>±0.157</sub> | **0.4341**<br><sub>±0.172</sub> | **0.2000**<br><sub>±0.163</sub> | **0.3481**<br><sub>±0.167</sub> | **0.8946**<br><sub>±0.031</sub> | **0.2342**<br><sub>±0.453</sub> | **0.3437**<br><sub>±0.190</sub> | **0.3257**<br><sub>±0.188</sub> | **0.2674**<br><sub>±0.179</sub> |
16
+
17
+ ## Cls head (nhãn CheXbert dương = positive|uncertain)
18
+
19
+ Precision 0.5556 · Recall 0.0144 · micro-F1 0.0281 · n = 6400
20
+
21
+ ## Route
22
+
23
+ ### MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉)
24
+
25
+ **Kết luận:** có phân hoá giữa các expert — cosine tham số ΔW (TB ngoài đường chéo) **0.109** (ngưỡng 0.9); hệ số biến thiên của route qua ảnh **0.073** (tĩnh nếu < 0.05); 6400 ảnh test.
26
+
27
+ ![E-Clinical | E-E — MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉)](e_heatmap.png)
28
+
29
+ ## Vector c hiệu dụng theo expert (val)
30
+
31
+ c_i = k_i·E/N tại token cuối chuỗi (k_i = số token expert i lấy); trung bình trên tập val. c cố định ⇒ ≈ c; Adaptive Capacity ⇒ c_i tự thay đổi theo độ chuyên biệt (H_i thấp ⇒ về sàn 1).
32
+
33
+ | Bước val | c_E0 | c_E1 | c_E2 | c_E3 |
34
+ |---|---|---|---|---|
35
+ | 40 | 1.385 | 2.147 | 3.069 | 2.598 |
36
+ | 80 | 1.074 | 2.030 | 3.128 | 1.612 |
37
+ | 120 | 1.016 | 1.974 | 3.130 | 1.460 |
38
+ | 160 | 1.005 | 1.939 | 3.112 | 1.427 |
39
+ | 200 | 1.004 | 1.848 | 3.101 | 1.317 |
40
+ | 240 | 1.004 | 1.824 | 3.096 | 1.273 |
41
+ | 280 | 1.004 | 1.730 | 3.076 | 1.226 |
42
+ | 320 | 1.004 | 1.694 | 3.058 | 1.218 |
43
+ | 360 | 1.004 | 1.682 | 3.059 | 1.211 |
44
+ | 400 | 1.004 | 1.682 | 3.059 | 1.210 |
3.1_c_vs_baseline_v4/arms/ac_e4/e_clinical_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,0.029001990028797858,0.027284941347483677,0.010297911329383999,0.01007343913995225,0.023650292934116662,0.026346032738128414,0.009358882635251205,0.004257906313219938,0.00224755178824077,0.033779593378869335,-0.01773659597821672,-0.010203219566820632,0.06695860859306262,0.008552539969362772
3
+ E1,0.013149081643942758,0.011316749892327347,-0.002630078240569749,-0.0060997687069583246,0.009321234257296055,-0.025659469048783287,-0.002031506161697511,-0.028878098161536195,-0.001894222704743286,-0.03871303728781846,-0.0005436320380950803,0.01407555604224104,-0.07017143874114694,-0.0017351947548875624
4
+ E2,-0.022981609038170536,-0.014948283062809558,-0.012578696277354069,0.00755789582553853,-0.03758045929393058,-0.018007723624421462,-0.008586793863235813,-0.011409530630742452,-0.020012122251719482,0.0037436990184401622,0.0007210645789809787,-0.002308193404422158,-0.06997667390181819,0.007707783526132676
5
+ E3,0.001494350211941346,-0.006833921743423601,0.005967846165088293,-0.009691384812355036,0.014237149645816086,0.034217474600459284,0.0013650611300677468,0.03511663309766913,0.023921815041397038,0.014029556168968818,0.00011523776591231422,-0.014578293074283013,0.11637725258145656,-0.0013850560003802852
3.1_c_vs_baseline_v4/arms/ac_e4/e_e_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,0.009273679000033248,-0.006155315776942644,-0.007147315642929719,0.0096762059761657
3
+ E1,-0.006155315776942644,0.008556202861925887,0.0047314162748929465,-0.010734094390851451
4
+ E2,-0.007147315642929719,0.0047314162748929465,0.00828392318632153,-0.010190944627756894
5
+ E3,0.0096762059761657,-0.010734094390851451,-0.010190944627756894,0.017314251894865813
3.1_c_vs_baseline_v4/arms/ac_e4/e_heatmap.png ADDED

Git LFS Details

  • SHA256: 33105d13e45811e310435c30407537065af67d4e09bcce87d93ddaa5e40f946a
  • Pointer size: 131 Bytes
  • Size of remote file: 133 kB
3.1_c_vs_baseline_v4/arms/ac_e4/route_verdict.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "route_corr_offdiag": -0.2766297042401296,
3
+ "param_cos_offdiag": 0.10896808910684253,
4
+ "route_cv": 0.07330342882834558,
5
+ "threshold": 0.9,
6
+ "cv_floor": 0.05,
7
+ "collapsed_by_param": false,
8
+ "collapsed_by_route": false,
9
+ "collapsed": false,
10
+ "reading": "có phân hoá giữa các expert",
11
+ "n_images": 6400,
12
+ "mean_weight": {
13
+ "E0": 0.0965817175000001,
14
+ "E1": 0.2386272445312495,
15
+ "E2": 0.4926115032812514,
16
+ "E3": 0.16430357109374988
17
+ },
18
+ "top1_share": {
19
+ "E0": 0.0,
20
+ "E1": 0.0,
21
+ "E2": 1.0,
22
+ "E3": 0.0
23
+ }
24
+ }
3.1_c_vs_baseline_v4/arms/ac_e4/summary.json ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm_id": "ac_e4",
3
+ "name": "MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉)",
4
+ "exp": "3.1",
5
+ "n_eval": 6400,
6
+ "setup": {
7
+ "decoder": "moe",
8
+ "E": 4,
9
+ "c": 1.0,
10
+ "adaptive": true,
11
+ "kind": "ft",
12
+ "params": 35381602,
13
+ "minutes": 610.1
14
+ },
15
+ "scores": {
16
+ "arm_id": "ac_e4",
17
+ "arm": "MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉)",
18
+ "n": 6400,
19
+ "bleu1_sd": 0.2128618864,
20
+ "bleu1": 0.3073806313,
21
+ "bleu1_sd_subset": 0.0032024716,
22
+ "bleu2_sd": 0.185561998,
23
+ "bleu2": 0.234734377,
24
+ "bleu2_sd_subset": 0.002705292,
25
+ "bleu3_sd": 0.1684718134,
26
+ "bleu3": 0.1815097373,
27
+ "bleu3_sd_subset": 0.0025268999,
28
+ "bleu4_sd": 0.1574260825,
29
+ "bleu4": 0.1456190378,
30
+ "bleu4_sd_subset": 0.0023905606,
31
+ "rouge1_sd": 0.1724715866,
32
+ "rouge1": 0.4341195798,
33
+ "rouge1_sd_subset": 0.0027289373,
34
+ "rouge2_sd": 0.163340908,
35
+ "rouge2": 0.2000067666,
36
+ "rouge2_sd_subset": 0.0025369064,
37
+ "rougeL_sd": 0.1666968352,
38
+ "rougeL": 0.3481042234,
39
+ "rougeL_sd_subset": 0.0020676272,
40
+ "bertscore_f1_sd": 0.0308231734,
41
+ "bertscore_f1": 0.894627884,
42
+ "bertscore_f1_sd_subset": 0.0005328467,
43
+ "chexbert_f1_sd": 0.4525653132,
44
+ "chexbert_f1": 0.234173922,
45
+ "chexbert_f1_sd_subset": 0.0056707215,
46
+ "radgraph_f1_entity_sd": 0.1900821399,
47
+ "radgraph_f1_entity": 0.3436502199,
48
+ "radgraph_f1_entity_sd_subset": 0.0035857742,
49
+ "radgraph_f1_partial_sd": 0.1881625291,
50
+ "radgraph_f1_partial": 0.325732399,
51
+ "radgraph_f1_partial_sd_subset": 0.003267391,
52
+ "radgraph_f1_complete_sd": 0.1786051742,
53
+ "radgraph_f1_complete": 0.2673729442,
54
+ "radgraph_f1_complete_sd_subset": 0.0027223142
55
+ },
56
+ "cls_head": {
57
+ "arm": "MoE-DoRA · c=a (Adaptive), E=4 (k_i=⌈max(N/E, e^H_i)⌉)",
58
+ "precision": 0.5555555556,
59
+ "recall": 0.0143921436,
60
+ "micro_F1": 0.0280574352,
61
+ "n": 6400
62
+ },
63
+ "route": {
64
+ "verdict": {
65
+ "route_corr_offdiag": -0.2766297042401296,
66
+ "param_cos_offdiag": 0.10896808910684253,
67
+ "route_cv": 0.07330342882834558,
68
+ "threshold": 0.9,
69
+ "cv_floor": 0.05,
70
+ "collapsed_by_param": false,
71
+ "collapsed_by_route": false,
72
+ "collapsed": false,
73
+ "reading": "có phân hoá giữa các expert",
74
+ "n_images": 6400,
75
+ "mean_weight": {
76
+ "E0": 0.0965817175000001,
77
+ "E1": 0.2386272445312495,
78
+ "E2": 0.4926115032812514,
79
+ "E3": 0.16430357109374988
80
+ },
81
+ "top1_share": {
82
+ "E0": 0.0,
83
+ "E1": 0.0,
84
+ "E2": 1.0,
85
+ "E3": 0.0
86
+ }
87
+ },
88
+ "heatmap": "e_heatmap.png"
89
+ },
90
+ "status": {
91
+ "route_verdict": {
92
+ "route_corr_offdiag": -0.2766297042401296,
93
+ "param_cos_offdiag": 0.10896808910684253,
94
+ "route_cv": 0.07330342882834558,
95
+ "threshold": 0.9,
96
+ "cv_floor": 0.05,
97
+ "collapsed_by_param": false,
98
+ "collapsed_by_route": false,
99
+ "collapsed": false,
100
+ "reading": "có phân hoá giữa các expert",
101
+ "n_images": 6400,
102
+ "mean_weight": {
103
+ "E0": 0.0965817175000001,
104
+ "E1": 0.2386272445312495,
105
+ "E2": 0.4926115032812514,
106
+ "E3": 0.16430357109374988
107
+ },
108
+ "top1_share": {
109
+ "E0": 0.0,
110
+ "E1": 0.0,
111
+ "E2": 1.0,
112
+ "E3": 0.0
113
+ }
114
+ }
115
+ },
116
+ "metrics": [
117
+ "bleu1",
118
+ "bleu2",
119
+ "bleu3",
120
+ "bleu4",
121
+ "rouge1",
122
+ "rouge2",
123
+ "rougeL",
124
+ "bertscore_f1",
125
+ "chexbert_f1",
126
+ "radgraph_f1_entity",
127
+ "radgraph_f1_partial",
128
+ "radgraph_f1_complete"
129
+ ],
130
+ "c_vec_val": [
131
+ {
132
+ "step": 40,
133
+ "c_e0": 1.3849223461002111,
134
+ "c_e1": 2.147181037813425,
135
+ "c_e2": 3.0694774836301804,
136
+ "c_e3": 2.597794570028782
137
+ },
138
+ {
139
+ "step": 80,
140
+ "c_e0": 1.0740269664674997,
141
+ "c_e1": 2.029694128781557,
142
+ "c_e2": 3.128379587084055,
143
+ "c_e3": 1.611822735518217
144
+ },
145
+ {
146
+ "step": 120,
147
+ "c_e0": 1.0156533643603325,
148
+ "c_e1": 1.9742577988654375,
149
+ "c_e2": 3.129681780934334,
150
+ "c_e3": 1.459740862250328
151
+ },
152
+ {
153
+ "step": 160,
154
+ "c_e0": 1.005408527329564,
155
+ "c_e1": 1.9388855788856745,
156
+ "c_e2": 3.112347334623337,
157
+ "c_e3": 1.4266904965043068
158
+ },
159
+ {
160
+ "step": 200,
161
+ "c_e0": 1.0038908869028091,
162
+ "c_e1": 1.8480207026004791,
163
+ "c_e2": 3.1006546318531036,
164
+ "c_e3": 1.3171690870076418
165
+ },
166
+ {
167
+ "step": 240,
168
+ "c_e0": 1.0038908869028091,
169
+ "c_e1": 1.8244619201868773,
170
+ "c_e2": 3.0961780175566673,
171
+ "c_e3": 1.2731978837400675
172
+ },
173
+ {
174
+ "step": 280,
175
+ "c_e0": 1.0038908869028091,
176
+ "c_e1": 1.729996770620346,
177
+ "c_e2": 3.0755213648080826,
178
+ "c_e3": 1.2255167737603188
179
+ },
180
+ {
181
+ "step": 320,
182
+ "c_e0": 1.0038908869028091,
183
+ "c_e1": 1.6936077792197466,
184
+ "c_e2": 3.05838755518198,
185
+ "c_e3": 1.2175792027264833
186
+ },
187
+ {
188
+ "step": 360,
189
+ "c_e0": 1.0038908869028091,
190
+ "c_e1": 1.6823519803583622,
191
+ "c_e2": 3.059245239943266,
192
+ "c_e3": 1.2107801083475351
193
+ },
194
+ {
195
+ "step": 400,
196
+ "c_e0": 1.0038908869028091,
197
+ "c_e1": 1.6815382223576307,
198
+ "c_e2": 3.058855816721916,
199
+ "c_e3": 1.2097816281020641
200
+ }
201
+ ]
202
+ }
3.1_c_vs_baseline_v4/arms/baseline_infer/REPORT.md ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # §3.1 — Baseline · Gemma 3 gốc
2
+
3
+ Greedy · QDoRA r=4 · 6400 ảnh test. Báo cáo con của riêng arm này; báo cáo cuối của mục chỉ đọc lại `summary.json` cùng thư mục.
4
+
5
+ ## Thiết lập
6
+
7
+ | Arm | Decoder | E | c | k mỗi expert | Cls head | Tham số học | Phút |
8
+ |---|---|---|---|---|---|---|---|
9
+ | Baseline · Gemma 3 gốc | — | — | — | — | — | 0 | — |
10
+
11
+ ## Điểm
12
+
13
+ | Mô hình | B-1 | B-2 | B-3 | B-4 | R-1 | R-2 | R-L | BS-F1 | CX-F1 | RG-E | RG-P | RG-C |
14
+ | :--- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
15
+ | Baseline · Gemma 3 gốc | **0.3233**<br><sub>±0.101</sub> | **0.1958**<br><sub>±0.074</sub> | **0.1177**<br><sub>±0.059</sub> | **0.0691**<br><sub>±0.044</sub> | **0.3291**<br><sub>±0.088</sub> | **0.0887**<br><sub>±0.054</sub> | **0.2104**<br><sub>±0.066</sub> | **0.8640**<br><sub>±0.020</sub> | **0.2144**<br><sub>±0.257</sub> | **0.2212**<br><sub>±0.110</sub> | **0.2027**<br><sub>±0.104</sub> | **0.1199**<br><sub>±0.084</sub> |
3.1_c_vs_baseline_v4/arms/baseline_infer/summary.json ADDED
@@ -0,0 +1,78 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm_id": "baseline_infer",
3
+ "name": "Baseline · Gemma 3 gốc",
4
+ "exp": "3.1",
5
+ "n_eval": 6400,
6
+ "setup": {
7
+ "decoder": "—",
8
+ "E": null,
9
+ "c": null,
10
+ "adaptive": false,
11
+ "kind": "infer",
12
+ "params": 0,
13
+ "minutes": null
14
+ },
15
+ "scores": {
16
+ "arm_id": "baseline_infer",
17
+ "arm": "Baseline · Gemma 3 gốc",
18
+ "n": 6400,
19
+ "bleu1_sd": 0.1011163448,
20
+ "bleu1": 0.3233420633,
21
+ "bleu1_sd_subset": 0.0005077294,
22
+ "bleu2_sd": 0.0743090859,
23
+ "bleu2": 0.1958227161,
24
+ "bleu2_sd_subset": 0.0005655452,
25
+ "bleu3_sd": 0.0590879176,
26
+ "bleu3": 0.1176898732,
27
+ "bleu3_sd_subset": 0.0006898889,
28
+ "bleu4_sd": 0.044375494,
29
+ "bleu4": 0.0691365918,
30
+ "bleu4_sd_subset": 0.0005623433,
31
+ "rouge1_sd": 0.0878925189,
32
+ "rouge1": 0.3291377718,
33
+ "rouge1_sd_subset": 0.0006640082,
34
+ "rouge2_sd": 0.0544602581,
35
+ "rouge2": 0.0887339278,
36
+ "rouge2_sd_subset": 0.0010074253,
37
+ "rougeL_sd": 0.0664438905,
38
+ "rougeL": 0.2103935063,
39
+ "rougeL_sd_subset": 0.0008653097,
40
+ "bertscore_f1_sd": 0.0204247638,
41
+ "bertscore_f1": 0.8639701964,
42
+ "bertscore_f1_sd_subset": 0.0002295768,
43
+ "chexbert_f1_sd": 0.2572580935,
44
+ "chexbert_f1": 0.2144093534,
45
+ "chexbert_f1_sd_subset": 0.0041883985,
46
+ "radgraph_f1_entity_sd": 0.109671111,
47
+ "radgraph_f1_entity": 0.2212124356,
48
+ "radgraph_f1_entity_sd_subset": 0.0017850696,
49
+ "radgraph_f1_partial_sd": 0.1040774069,
50
+ "radgraph_f1_partial": 0.2026760622,
51
+ "radgraph_f1_partial_sd_subset": 0.0014018197,
52
+ "radgraph_f1_complete_sd": 0.0839295975,
53
+ "radgraph_f1_complete": 0.1198624362,
54
+ "radgraph_f1_complete_sd_subset": 0.0012224729
55
+ },
56
+ "cls_head": null,
57
+ "route": null,
58
+ "status": {
59
+ "infer": {
60
+ "ok": true,
61
+ "minutes": 67.6
62
+ }
63
+ },
64
+ "metrics": [
65
+ "bleu1",
66
+ "bleu2",
67
+ "bleu3",
68
+ "bleu4",
69
+ "rouge1",
70
+ "rouge2",
71
+ "rougeL",
72
+ "bertscore_f1",
73
+ "chexbert_f1",
74
+ "radgraph_f1_entity",
75
+ "radgraph_f1_partial",
76
+ "radgraph_f1_complete"
77
+ ]
78
+ }
3.1_c_vs_baseline_v4/arms/c1_e4/REPORT.md ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # §3.1 — MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉)
2
+
3
+ Greedy · QDoRA r=4 · 6400 ảnh test. Báo cáo con của riêng arm này; báo cáo cuối của mục chỉ đọc lại `summary.json` cùng thư mục.
4
+
5
+ ## Thiết lập
6
+
7
+ | Arm | Decoder | E | c | k mỗi expert | Cls head | Tham số học | Phút |
8
+ |---|---|---|---|---|---|---|---|
9
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) | moe | 4 | 1 | ⌈c·N/E⌉ | đồng huấn luyện | 35,381,602 | 608.6 |
10
+
11
+ ## Điểm
12
+
13
+ | Mô hình | B-1 | B-2 | B-3 | B-4 | R-1 | R-2 | R-L | BS-F1 | CX-F1 | RG-E | RG-P | RG-C |
14
+ | :--- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
15
+ | MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉) | **0.3035**<br><sub>±0.215</sub> | **0.2323**<br><sub>±0.187</sub> | **0.1799**<br><sub>±0.169</sub> | **0.1444**<br><sub>±0.158</sub> | **0.4312**<br><sub>±0.175</sub> | **0.1986**<br><sub>±0.164</sub> | **0.3467**<br><sub>±0.168</sub> | **0.8946**<br><sub>±0.031</sub> | **0.2254**<br><sub>±0.454</sub> | **0.3413**<br><sub>±0.191</sub> | **0.3237**<br><sub>±0.189</sub> | **0.2659**<br><sub>±0.179</sub> |
16
+
17
+ ## Cls head (nhãn CheXbert dương = positive|uncertain)
18
+
19
+ Precision 0.6009 · Recall 0.0113 · micro-F1 0.0223 · n = 6400
20
+
21
+ ## Route
22
+
23
+ ### MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉)
24
+
25
+ **Kết luận:** có phân hoá giữa các expert — cosine tham số ΔW (TB ngoài đường chéo) **0.073** (ngưỡng 0.9); hệ số biến thiên của route qua ảnh **0.124** (tĩnh nếu < 0.05); 6400 ảnh test.
26
+
27
+ ![E-Clinical | E-E — MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉)](e_heatmap.png)
28
+
29
+ ## Vector c hiệu dụng theo expert (val)
30
+
31
+ c_i = k_i·E/N tại token cuối chuỗi (k_i = số token expert i lấy); trung bình trên tập val. c cố định ⇒ ≈ c; Adaptive Capacity ⇒ c_i tự thay đổi theo độ chuyên biệt (H_i thấp ⇒ về sàn 1).
32
+
33
+ | Bước val | c_E0 | c_E1 | c_E2 | c_E3 |
34
+ |---|---|---|---|---|
35
+ | 40 | 1.004 | 1.004 | 1.004 | 1.004 |
36
+ | 80 | 1.004 | 1.004 | 1.004 | 1.004 |
37
+ | 120 | 1.004 | 1.004 | 1.004 | 1.004 |
38
+ | 160 | 1.004 | 1.004 | 1.004 | 1.004 |
39
+ | 200 | 1.004 | 1.004 | 1.004 | 1.004 |
40
+ | 240 | 1.004 | 1.004 | 1.004 | 1.004 |
41
+ | 280 | 1.004 | 1.004 | 1.004 | 1.004 |
42
+ | 320 | 1.004 | 1.004 | 1.004 | 1.004 |
43
+ | 360 | 1.004 | 1.004 | 1.004 | 1.004 |
44
+ | 400 | 1.004 | 1.004 | 1.004 | 1.004 |
3.1_c_vs_baseline_v4/arms/c1_e4/e_clinical_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,0.020802602671289583,-0.0048122384070191475,-0.016545230157883884,0.024412019598496596,-0.03233482815187352,0.005184547594760089,-0.013824617721318417,-0.03728982823402704,0.006609487291886123,-0.0008012556369530595,-0.013457426744585365,-0.01163733714032374,-0.030492927953746807,0.03066511417163733
3
+ E1,-0.020916341474653172,-0.015132093344330223,0.0009880655125392567,-0.02141402355167992,-0.005207085955001309,-0.03360727902531349,0.011275043893673039,-0.031038008179371582,-0.03175570787506757,-0.05415911557123908,-0.00012395535510693724,0.012586149462917886,-0.1078066856476099,0.038248511667539814
4
+ E2,0.04223480751727722,0.09035062805266311,0.06123167250990046,-0.010578698575576069,0.09558577473535945,0.0315623671552873,0.02956957029452947,0.07907448218083776,0.014366820039710189,0.06740220261706717,-0.010172372284838313,0.010108766371818683,0.09696745783769822,-0.09898284792035915
5
+ E3,0.002045298059343228,-0.0087277243050781,-0.009631795907836832,-0.0011572155690196082,-0.016362141240546378,-0.007443036493606882,0.01003872897362044,0.010934751536764955,0.00397382713898063,-0.00026755228783946326,-0.006018877491987675,0.00938983167566698,-0.013897710359303667,-0.005048718337458104
3.1_c_vs_baseline_v4/arms/c1_e4/e_e_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,0.006210694413990362,0.004243199759873831,-0.013044045509177296,0.0004267670354173167
3
+ E1,0.004243199759873831,0.020558349202412764,-0.02386079095159684,0.0015260725927867503
4
+ E2,-0.013044045509177296,-0.02386079095159684,0.05522361692229719,-0.002569794421163723
5
+ E3,0.0004267670354173167,0.0015260725927867503,-0.002569794421163723,0.0010768250063779745
3.1_c_vs_baseline_v4/arms/c1_e4/e_heatmap.png ADDED

Git LFS Details

  • SHA256: 9f74e19ced7e26ac26d48a27580899b18e5c0fc1e9bb0a98a1cba90496068a0c
  • Pointer size: 131 Bytes
  • Size of remote file: 137 kB
3.1_c_vs_baseline_v4/arms/c1_e4/route_verdict.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "route_corr_offdiag": -0.13875327534314474,
3
+ "param_cos_offdiag": 0.07319301052722915,
4
+ "route_cv": 0.12412635149631922,
5
+ "threshold": 0.9,
6
+ "cv_floor": 0.05,
7
+ "collapsed_by_param": false,
8
+ "collapsed_by_route": false,
9
+ "collapsed": false,
10
+ "reading": "có phân hoá giữa các expert",
11
+ "n_images": 6400,
12
+ "mean_weight": {
13
+ "E0": 0.12108115031250062,
14
+ "E1": 0.16905713812499987,
15
+ "E2": 0.1870031420312501,
16
+ "E3": 0.13392153968749995
17
+ },
18
+ "top1_share": {
19
+ "E0": 0.00421875,
20
+ "E1": 0.30234375,
21
+ "E2": 0.68234375,
22
+ "E3": 0.01109375
23
+ }
24
+ }
3.1_c_vs_baseline_v4/arms/c1_e4/summary.json ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm_id": "c1_e4",
3
+ "name": "MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉)",
4
+ "exp": "3.1",
5
+ "n_eval": 6400,
6
+ "setup": {
7
+ "decoder": "moe",
8
+ "E": 4,
9
+ "c": 1.0,
10
+ "adaptive": false,
11
+ "kind": "ft",
12
+ "params": 35381602,
13
+ "minutes": 608.6
14
+ },
15
+ "scores": {
16
+ "arm_id": "c1_e4",
17
+ "arm": "MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉)",
18
+ "n": 6400,
19
+ "bleu1_sd": 0.215023243,
20
+ "bleu1": 0.3034763175,
21
+ "bleu1_sd_subset": 0.0032694719,
22
+ "bleu2_sd": 0.1867615392,
23
+ "bleu2": 0.2323361527,
24
+ "bleu2_sd_subset": 0.0028228047,
25
+ "bleu3_sd": 0.168985995,
26
+ "bleu3": 0.1798820572,
27
+ "bleu3_sd_subset": 0.0025798371,
28
+ "bleu4_sd": 0.1575156628,
29
+ "bleu4": 0.1444277763,
30
+ "bleu4_sd_subset": 0.0024114239,
31
+ "rouge1_sd": 0.1747194218,
32
+ "rouge1": 0.4312273672,
33
+ "rouge1_sd_subset": 0.0028065937,
34
+ "rouge2_sd": 0.1636255444,
35
+ "rouge2": 0.1986012589,
36
+ "rouge2_sd_subset": 0.0026312504,
37
+ "rougeL_sd": 0.1677391685,
38
+ "rougeL": 0.3466607894,
39
+ "rougeL_sd_subset": 0.0023632266,
40
+ "bertscore_f1_sd": 0.0306730559,
41
+ "bertscore_f1": 0.894598752,
42
+ "bertscore_f1_sd_subset": 0.0005386786,
43
+ "chexbert_f1_sd": 0.4540392366,
44
+ "chexbert_f1": 0.2253746053,
45
+ "chexbert_f1_sd_subset": 0.0055399017,
46
+ "radgraph_f1_entity_sd": 0.1914962061,
47
+ "radgraph_f1_entity": 0.3413127098,
48
+ "radgraph_f1_entity_sd_subset": 0.0035337238,
49
+ "radgraph_f1_partial_sd": 0.189152867,
50
+ "radgraph_f1_partial": 0.3236600279,
51
+ "radgraph_f1_partial_sd_subset": 0.0032816341,
52
+ "radgraph_f1_complete_sd": 0.179042327,
53
+ "radgraph_f1_complete": 0.2659483849,
54
+ "radgraph_f1_complete_sd_subset": 0.0027603427
55
+ },
56
+ "cls_head": {
57
+ "arm": "MoE-DoRA · c=1, E=4 (k=⌈c·N/E⌉)",
58
+ "precision": 0.600896861,
59
+ "recall": 0.0113443955,
60
+ "micro_F1": 0.0222683839,
61
+ "n": 6400
62
+ },
63
+ "route": {
64
+ "verdict": {
65
+ "route_corr_offdiag": -0.13875327534314474,
66
+ "param_cos_offdiag": 0.07319301052722915,
67
+ "route_cv": 0.12412635149631922,
68
+ "threshold": 0.9,
69
+ "cv_floor": 0.05,
70
+ "collapsed_by_param": false,
71
+ "collapsed_by_route": false,
72
+ "collapsed": false,
73
+ "reading": "có phân hoá giữa các expert",
74
+ "n_images": 6400,
75
+ "mean_weight": {
76
+ "E0": 0.12108115031250062,
77
+ "E1": 0.16905713812499987,
78
+ "E2": 0.1870031420312501,
79
+ "E3": 0.13392153968749995
80
+ },
81
+ "top1_share": {
82
+ "E0": 0.00421875,
83
+ "E1": 0.30234375,
84
+ "E2": 0.68234375,
85
+ "E3": 0.01109375
86
+ }
87
+ },
88
+ "heatmap": "e_heatmap.png"
89
+ },
90
+ "status": {
91
+ "route_verdict": {
92
+ "route_corr_offdiag": -0.13875327534314474,
93
+ "param_cos_offdiag": 0.07319301052722915,
94
+ "route_cv": 0.12412635149631922,
95
+ "threshold": 0.9,
96
+ "cv_floor": 0.05,
97
+ "collapsed_by_param": false,
98
+ "collapsed_by_route": false,
99
+ "collapsed": false,
100
+ "reading": "có phân hoá giữa các expert",
101
+ "n_images": 6400,
102
+ "mean_weight": {
103
+ "E0": 0.12108115031250062,
104
+ "E1": 0.16905713812499987,
105
+ "E2": 0.1870031420312501,
106
+ "E3": 0.13392153968749995
107
+ },
108
+ "top1_share": {
109
+ "E0": 0.00421875,
110
+ "E1": 0.30234375,
111
+ "E2": 0.68234375,
112
+ "E3": 0.01109375
113
+ }
114
+ }
115
+ },
116
+ "metrics": [
117
+ "bleu1",
118
+ "bleu2",
119
+ "bleu3",
120
+ "bleu4",
121
+ "rouge1",
122
+ "rouge2",
123
+ "rougeL",
124
+ "bertscore_f1",
125
+ "chexbert_f1",
126
+ "radgraph_f1_entity",
127
+ "radgraph_f1_partial",
128
+ "radgraph_f1_complete"
129
+ ],
130
+ "c_vec_val": [
131
+ {
132
+ "step": 40,
133
+ "c_e0": 1.0038908869028091,
134
+ "c_e1": 1.0038908869028091,
135
+ "c_e2": 1.0038908869028091,
136
+ "c_e3": 1.0038908869028091
137
+ },
138
+ {
139
+ "step": 80,
140
+ "c_e0": 1.0038908869028091,
141
+ "c_e1": 1.0038908869028091,
142
+ "c_e2": 1.0038908869028091,
143
+ "c_e3": 1.0038908869028091
144
+ },
145
+ {
146
+ "step": 120,
147
+ "c_e0": 1.0038908869028091,
148
+ "c_e1": 1.0038908869028091,
149
+ "c_e2": 1.0038908869028091,
150
+ "c_e3": 1.0038908869028091
151
+ },
152
+ {
153
+ "step": 160,
154
+ "c_e0": 1.0038908869028091,
155
+ "c_e1": 1.0038908869028091,
156
+ "c_e2": 1.0038908869028091,
157
+ "c_e3": 1.0038908869028091
158
+ },
159
+ {
160
+ "step": 200,
161
+ "c_e0": 1.0038908869028091,
162
+ "c_e1": 1.0038908869028091,
163
+ "c_e2": 1.0038908869028091,
164
+ "c_e3": 1.0038908869028091
165
+ },
166
+ {
167
+ "step": 240,
168
+ "c_e0": 1.0038908869028091,
169
+ "c_e1": 1.0038908869028091,
170
+ "c_e2": 1.0038908869028091,
171
+ "c_e3": 1.0038908869028091
172
+ },
173
+ {
174
+ "step": 280,
175
+ "c_e0": 1.0038908869028091,
176
+ "c_e1": 1.0038908869028091,
177
+ "c_e2": 1.0038908869028091,
178
+ "c_e3": 1.0038908869028091
179
+ },
180
+ {
181
+ "step": 320,
182
+ "c_e0": 1.0038908869028091,
183
+ "c_e1": 1.0038908869028091,
184
+ "c_e2": 1.0038908869028091,
185
+ "c_e3": 1.0038908869028091
186
+ },
187
+ {
188
+ "step": 360,
189
+ "c_e0": 1.0038908869028091,
190
+ "c_e1": 1.0038908869028091,
191
+ "c_e2": 1.0038908869028091,
192
+ "c_e3": 1.0038908869028091
193
+ },
194
+ {
195
+ "step": 400,
196
+ "c_e0": 1.0038908869028091,
197
+ "c_e1": 1.0038908869028091,
198
+ "c_e2": 1.0038908869028091,
199
+ "c_e3": 1.0038908869028091
200
+ }
201
+ ]
202
+ }
3.1_c_vs_baseline_v4/arms/c2_e4/REPORT.md ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # §3.1 — MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉)
2
+
3
+ Greedy · QDoRA r=4 · 6400 ảnh test. Báo cáo con của riêng arm này; báo cáo cuối của mục chỉ đọc lại `summary.json` cùng thư mục.
4
+
5
+ ## Thiết lập
6
+
7
+ | Arm | Decoder | E | c | k mỗi expert | Cls head | Tham số học | Phút |
8
+ |---|---|---|---|---|---|---|---|
9
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | moe | 4 | 2 | ⌈c·N/E⌉ | đồng huấn luyện | 35,381,602 | 607.5 |
10
+
11
+ ## Điểm
12
+
13
+ | Mô hình | B-1 | B-2 | B-3 | B-4 | R-1 | R-2 | R-L | BS-F1 | CX-F1 | RG-E | RG-P | RG-C |
14
+ | :--- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
15
+ | MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉) | **0.3050**<br><sub>±0.214</sub> | **0.2334**<br><sub>±0.186</sub> | **0.1807**<br><sub>±0.169</sub> | **0.1450**<br><sub>±0.157</sub> | **0.4323**<br><sub>±0.174</sub> | **0.1994**<br><sub>±0.164</sub> | **0.3471**<br><sub>±0.167</sub> | **0.8944**<br><sub>±0.031</sub> | **0.2306**<br><sub>±0.453</sub> | **0.3420**<br><sub>±0.191</sub> | **0.3242**<br><sub>±0.189</sub> | **0.2660**<br><sub>±0.179</sub> |
16
+
17
+ ## Cls head (nhãn CheXbert dương = positive|uncertain)
18
+
19
+ Precision 0.5633 · Recall 0.0151 · micro-F1 0.0294 · n = 6400
20
+
21
+ ## Route
22
+
23
+ ### MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉)
24
+
25
+ **Kết luận:** có phân hoá giữa các expert — cosine tham số ΔW (TB ngoài đường chéo) **0.108** (ngưỡng 0.9); hệ số biến thiên của route qua ảnh **0.099** (tĩnh nếu < 0.05); 6400 ảnh test.
26
+
27
+ ![E-Clinical | E-E — MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉)](e_heatmap.png)
28
+
29
+ ## Vector c hiệu dụng theo expert (val)
30
+
31
+ c_i = k_i·E/N tại token cuối chuỗi (k_i = số token expert i lấy); trung bình trên tập val. c cố định ⇒ ≈ c; Adaptive Capacity ⇒ c_i tự thay đổi theo độ chuyên biệt (H_i thấp ⇒ về sàn 1).
32
+
33
+ | Bước val | c_E0 | c_E1 | c_E2 | c_E3 |
34
+ |---|---|---|---|---|
35
+ | 40 | 2.003 | 2.003 | 2.003 | 2.003 |
36
+ | 80 | 2.003 | 2.003 | 2.003 | 2.003 |
37
+ | 120 | 2.003 | 2.003 | 2.003 | 2.003 |
38
+ | 160 | 2.003 | 2.003 | 2.003 | 2.003 |
39
+ | 200 | 2.003 | 2.003 | 2.003 | 2.003 |
40
+ | 240 | 2.003 | 2.003 | 2.003 | 2.003 |
41
+ | 280 | 2.003 | 2.003 | 2.003 | 2.003 |
42
+ | 320 | 2.003 | 2.003 | 2.003 | 2.003 |
43
+ | 360 | 2.003 | 2.003 | 2.003 | 2.003 |
44
+ | 400 | 2.003 | 2.003 | 2.003 | 2.003 |
3.1_c_vs_baseline_v4/arms/c2_e4/e_clinical_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,-0.007930492975845953,-0.04881091065676136,-0.03791276661676145,0.004952127864247708,-0.06455421957073118,-0.01382495857331236,-0.03736636727923224,-0.05092123453155257,0.022159097189434967,-0.029115130219275108,0.012191628462464682,-0.012000258177457993,-0.00828926244453073,0.06472117573409376
3
+ E1,-0.015576385135676975,-0.023929937884146435,-0.03351492647449298,-0.00350610730974592,-0.04526702650645237,-0.03964694784142206,-0.014502180767403516,-0.04288372521222152,-0.03399313803968396,-0.07826525900090851,0.0027131046128350154,0.008391644534461541,-0.18220310235068882,0.04084404019741215
4
+ E2,0.049256177805843356,0.10508498675236425,0.0739901963255877,-0.029963804525854735,0.11258389313152836,0.04480839435558625,0.0526032490597194,0.08097728952900618,0.0035921344565115177,0.10609956779355957,-0.013481732872184085,0.013406455311193798,0.13297483343745184,-0.11905355608961328
5
+ E3,-0.0007437161481991803,-0.024406732305714187,-0.007048606567558903,0.00765275786416813,-0.012906889465381516,0.017618747119828606,-0.0037002439072465897,0.017579284538730894,0.017057222349827074,0.012222427059557905,0.008925871365510743,-0.0019275822565562185,0.024445868530364845,0.009911620900084201
3.1_c_vs_baseline_v4/arms/c2_e4/e_e_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,0.01814379429181335,0.014352357246028246,-0.03459207163519112,0.0019279850290365465
3
+ E1,0.014352357246028246,0.04984609104381689,-0.05420610252216201,-0.0055870996434518785
4
+ E2,-0.03459207163519112,-0.05420610252216201,0.08733632008171109,0.0004955499734603486
5
+ E3,0.0019279850290365465,-0.0055870996434518785,0.0004955499734603486,0.0027237843762124864
3.1_c_vs_baseline_v4/arms/c2_e4/e_heatmap.png ADDED

Git LFS Details

  • SHA256: 7ef79585a6a138a10ab99b0b4ab9b152e4b80764d8e8c1a05e1bbdb7e818f949
  • Pointer size: 131 Bytes
  • Size of remote file: 140 kB
3.1_c_vs_baseline_v4/arms/c2_e4/route_verdict.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "route_corr_offdiag": -0.2659895499372867,
3
+ "param_cos_offdiag": 0.10780160052949572,
4
+ "route_cv": 0.09943060071331047,
5
+ "threshold": 0.9,
6
+ "cv_floor": 0.05,
7
+ "collapsed_by_param": false,
8
+ "collapsed_by_route": false,
9
+ "collapsed": false,
10
+ "reading": "có phân hoá giữa các expert",
11
+ "n_images": 6400,
12
+ "mean_weight": {
13
+ "E0": 0.1591673885937501,
14
+ "E1": 0.28544982921874923,
15
+ "E2": 0.2872016970312495,
16
+ "E3": 0.2024046340625001
17
+ },
18
+ "top1_share": {
19
+ "E0": 0.00015625,
20
+ "E1": 0.48453125,
21
+ "E2": 0.514375,
22
+ "E3": 0.0009375
23
+ }
24
+ }
3.1_c_vs_baseline_v4/arms/c2_e4/summary.json ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm_id": "c2_e4",
3
+ "name": "MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉)",
4
+ "exp": "3.1",
5
+ "n_eval": 6400,
6
+ "setup": {
7
+ "decoder": "moe",
8
+ "E": 4,
9
+ "c": 2.0,
10
+ "adaptive": false,
11
+ "kind": "ft",
12
+ "params": 35381602,
13
+ "minutes": 607.5
14
+ },
15
+ "scores": {
16
+ "arm_id": "c2_e4",
17
+ "arm": "MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉)",
18
+ "n": 6400,
19
+ "bleu1_sd": 0.2141265766,
20
+ "bleu1": 0.3050211868,
21
+ "bleu1_sd_subset": 0.0032340189,
22
+ "bleu2_sd": 0.1862535249,
23
+ "bleu2": 0.2334375692,
24
+ "bleu2_sd_subset": 0.0027522707,
25
+ "bleu3_sd": 0.1687741494,
26
+ "bleu3": 0.1807117588,
27
+ "bleu3_sd_subset": 0.0025254553,
28
+ "bleu4_sd": 0.1574946609,
29
+ "bleu4": 0.1450232277,
30
+ "bleu4_sd_subset": 0.0023262225,
31
+ "rouge1_sd": 0.1739281046,
32
+ "rouge1": 0.4323103709,
33
+ "rouge1_sd_subset": 0.0028094503,
34
+ "rouge2_sd": 0.1635966759,
35
+ "rouge2": 0.1993654915,
36
+ "rouge2_sd_subset": 0.0025566496,
37
+ "rougeL_sd": 0.1672368477,
38
+ "rougeL": 0.3471145767,
39
+ "rougeL_sd_subset": 0.0022345738,
40
+ "bertscore_f1_sd": 0.0308979411,
41
+ "bertscore_f1": 0.8944066738,
42
+ "bertscore_f1_sd_subset": 0.0005885396,
43
+ "chexbert_f1_sd": 0.4528986081,
44
+ "chexbert_f1": 0.2305590884,
45
+ "chexbert_f1_sd_subset": 0.005472442,
46
+ "radgraph_f1_entity_sd": 0.1911805327,
47
+ "radgraph_f1_entity": 0.3419757157,
48
+ "radgraph_f1_entity_sd_subset": 0.0038615329,
49
+ "radgraph_f1_partial_sd": 0.1892030596,
50
+ "radgraph_f1_partial": 0.3241758676,
51
+ "radgraph_f1_partial_sd_subset": 0.0035207155,
52
+ "radgraph_f1_complete_sd": 0.1791967734,
53
+ "radgraph_f1_complete": 0.2659690789,
54
+ "radgraph_f1_complete_sd_subset": 0.0029707607
55
+ },
56
+ "cls_head": {
57
+ "arm": "MoE-DoRA · c=2, E=4 (k=⌈c·N/E⌉)",
58
+ "precision": 0.5632911392,
59
+ "recall": 0.0150694209,
60
+ "micro_F1": 0.029353562,
61
+ "n": 6400
62
+ },
63
+ "route": {
64
+ "verdict": {
65
+ "route_corr_offdiag": -0.2659895499372867,
66
+ "param_cos_offdiag": 0.10780160052949572,
67
+ "route_cv": 0.09943060071331047,
68
+ "threshold": 0.9,
69
+ "cv_floor": 0.05,
70
+ "collapsed_by_param": false,
71
+ "collapsed_by_route": false,
72
+ "collapsed": false,
73
+ "reading": "có phân hoá giữa các expert",
74
+ "n_images": 6400,
75
+ "mean_weight": {
76
+ "E0": 0.1591673885937501,
77
+ "E1": 0.28544982921874923,
78
+ "E2": 0.2872016970312495,
79
+ "E3": 0.2024046340625001
80
+ },
81
+ "top1_share": {
82
+ "E0": 0.00015625,
83
+ "E1": 0.48453125,
84
+ "E2": 0.514375,
85
+ "E3": 0.0009375
86
+ }
87
+ },
88
+ "heatmap": "e_heatmap.png"
89
+ },
90
+ "status": {
91
+ "route_verdict": {
92
+ "route_corr_offdiag": -0.2659895499372867,
93
+ "param_cos_offdiag": 0.10780160052949572,
94
+ "route_cv": 0.09943060071331047,
95
+ "threshold": 0.9,
96
+ "cv_floor": 0.05,
97
+ "collapsed_by_param": false,
98
+ "collapsed_by_route": false,
99
+ "collapsed": false,
100
+ "reading": "có phân hoá giữa các expert",
101
+ "n_images": 6400,
102
+ "mean_weight": {
103
+ "E0": 0.1591673885937501,
104
+ "E1": 0.28544982921874923,
105
+ "E2": 0.2872016970312495,
106
+ "E3": 0.2024046340625001
107
+ },
108
+ "top1_share": {
109
+ "E0": 0.00015625,
110
+ "E1": 0.48453125,
111
+ "E2": 0.514375,
112
+ "E3": 0.0009375
113
+ }
114
+ }
115
+ },
116
+ "metrics": [
117
+ "bleu1",
118
+ "bleu2",
119
+ "bleu3",
120
+ "bleu4",
121
+ "rouge1",
122
+ "rouge2",
123
+ "rougeL",
124
+ "bertscore_f1",
125
+ "chexbert_f1",
126
+ "radgraph_f1_entity",
127
+ "radgraph_f1_partial",
128
+ "radgraph_f1_complete"
129
+ ],
130
+ "c_vec_val": [
131
+ {
132
+ "step": 40,
133
+ "c_e0": 2.0025149509310722,
134
+ "c_e1": 2.0025149509310722,
135
+ "c_e2": 2.0025149509310722,
136
+ "c_e3": 2.0025149509310722
137
+ },
138
+ {
139
+ "step": 80,
140
+ "c_e0": 2.0025149509310722,
141
+ "c_e1": 2.0025149509310722,
142
+ "c_e2": 2.0025149509310722,
143
+ "c_e3": 2.0025149509310722
144
+ },
145
+ {
146
+ "step": 120,
147
+ "c_e0": 2.0025149509310722,
148
+ "c_e1": 2.0025149509310722,
149
+ "c_e2": 2.0025149509310722,
150
+ "c_e3": 2.0025149509310722
151
+ },
152
+ {
153
+ "step": 160,
154
+ "c_e0": 2.0025149509310722,
155
+ "c_e1": 2.0025149509310722,
156
+ "c_e2": 2.0025149509310722,
157
+ "c_e3": 2.0025149509310722
158
+ },
159
+ {
160
+ "step": 200,
161
+ "c_e0": 2.0025149509310722,
162
+ "c_e1": 2.0025149509310722,
163
+ "c_e2": 2.0025149509310722,
164
+ "c_e3": 2.0025149509310722
165
+ },
166
+ {
167
+ "step": 240,
168
+ "c_e0": 2.0025149509310722,
169
+ "c_e1": 2.0025149509310722,
170
+ "c_e2": 2.0025149509310722,
171
+ "c_e3": 2.0025149509310722
172
+ },
173
+ {
174
+ "step": 280,
175
+ "c_e0": 2.0025149509310722,
176
+ "c_e1": 2.0025149509310722,
177
+ "c_e2": 2.0025149509310722,
178
+ "c_e3": 2.0025149509310722
179
+ },
180
+ {
181
+ "step": 320,
182
+ "c_e0": 2.0025149509310722,
183
+ "c_e1": 2.0025149509310722,
184
+ "c_e2": 2.0025149509310722,
185
+ "c_e3": 2.0025149509310722
186
+ },
187
+ {
188
+ "step": 360,
189
+ "c_e0": 2.0025149509310722,
190
+ "c_e1": 2.0025149509310722,
191
+ "c_e2": 2.0025149509310722,
192
+ "c_e3": 2.0025149509310722
193
+ },
194
+ {
195
+ "step": 400,
196
+ "c_e0": 2.0025149509310722,
197
+ "c_e1": 2.0025149509310722,
198
+ "c_e2": 2.0025149509310722,
199
+ "c_e3": 2.0025149509310722
200
+ }
201
+ ]
202
+ }
3.1_c_vs_baseline_v4/best.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "ac_e4",
3
+ "rule": "hạng lâm sàng TB; hoà ⇒ hạng văn bản TB",
4
+ "clinical_rank": 1.0,
5
+ "text_rank": 1.125,
6
+ "c": "a",
7
+ "E": 4,
8
+ "table": [
9
+ {
10
+ "arm_id": "ac_e4",
11
+ "clinical_rank": 1.0,
12
+ "text_rank": 1.125,
13
+ "final_rank": 1,
14
+ "tied_on_clinical": false
15
+ },
16
+ {
17
+ "arm_id": "c2_e4",
18
+ "clinical_rank": 2.0,
19
+ "text_rank": 2.25,
20
+ "final_rank": 2,
21
+ "tied_on_clinical": false
22
+ },
23
+ {
24
+ "arm_id": "c1_e4",
25
+ "clinical_rank": 3.0,
26
+ "text_rank": 3.0,
27
+ "final_rank": 3,
28
+ "tied_on_clinical": false
29
+ },
30
+ {
31
+ "arm_id": "baseline_infer",
32
+ "clinical_rank": 4.0,
33
+ "text_rank": 3.625,
34
+ "final_rank": 4,
35
+ "tied_on_clinical": false
36
+ }
37
+ ],
38
+ "experiment": "3.1",
39
+ "note": "GỢI Ý theo hạng trung bình — không qua kiểm định; người dùng quyết định"
40
+ }
3.1_c_vs_baseline_v4/config.resolved.yaml ADDED
@@ -0,0 +1,252 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ seed: 42
2
+ experiment: '3.1'
3
+ data:
4
+ repo: RGExp123/MIMIC_CXR_cleaned
5
+ repo_type: dataset
6
+ split_csv: test_split.csv
7
+ meta_csv: csv_helper/clean_dataset.csv
8
+ image_zip: img_dir.zip
9
+ key_column: img_rel
10
+ ref_column: text_gt
11
+ n_test: 6400
12
+ manifest_dir: ./data/exp30_v4
13
+ split_local_dir: ./data/splits_exp30_v4
14
+ split_hub_dir: split_ref_exp30_v4
15
+ split_seed: 42
16
+ split_patient_col: subject_id
17
+ reuse_manifests: true
18
+ cache_dir: ./data
19
+ image_max_side: 512
20
+ image_max_side_enforce: true
21
+ zip_mode: auto
22
+ zip_stream_min_rate: 1.0
23
+ train_split_csv: train_split.csv
24
+ n_train: 12800
25
+ val_split_csv: val_split.csv
26
+ n_val: 128
27
+ assert_no_leak: true
28
+ split_balance:
29
+ max_abs_rate_diff: 0.002
30
+ max_js: 0.0001
31
+ enforce: true
32
+ image_policy:
33
+ unified: true
34
+ resize_to: 512
35
+ interpolation: bicubic
36
+ keep_aspect: true
37
+ pad_to_square: true
38
+ pad_fill: 0
39
+ allow_native_match: false
40
+ per_model_normalize_only: true
41
+ base:
42
+ hf_id: google/gemma-3-4b-it
43
+ runner: gemma3_ec
44
+ dtype: bfloat16
45
+ attn_implementation_train: eager
46
+ attn_implementation_infer: sdpa
47
+ disable_compile: auto
48
+ batch_size: 512
49
+ infer_batch_probe: true
50
+ infer_batch_probe_sizes: pow2
51
+ device_map_train: single
52
+ device_map_infer: auto
53
+ image:
54
+ do_pan_and_scan: false
55
+ processor_size: 896
56
+ prompt:
57
+ system: You are an experienced radiologist reading a chest X-ray.
58
+ user: 'Write the radiology report for this chest X-ray.
59
+
60
+ Answer in exactly this format, with no preamble and no extra sections:
61
+
62
+ Findings: <the radiological findings>
63
+
64
+ Impression: <the overall impression>
65
+
66
+ '
67
+ generate:
68
+ max_new_tokens: 256
69
+ do_sample: false
70
+ num_beams: 1
71
+ length_penalty: 1.0
72
+ early_stopping: false
73
+ no_repeat_ngram_size: 0
74
+ repetition_penalty: 1.0
75
+ num_return_sequences: 1
76
+ lora:
77
+ rank: 4
78
+ heads:
79
+ d_c: 256
80
+ n_classes: 4
81
+ classifier:
82
+ gamma: 0.1
83
+ labels:
84
+ cache_dir: ./data
85
+ batch_size: 32
86
+ device: auto
87
+ train:
88
+ n_epochs: 1
89
+ steps_per_epoch: 400
90
+ epoch_split: false
91
+ batch_probe: true
92
+ batch_probe_sizes: divisors
93
+ per_device_batch_size: 8
94
+ effective_batch_size: 32
95
+ grad_accum: 1
96
+ max_seq_len: 1024
97
+ gradient_checkpointing: true
98
+ weight_decay: 0.001
99
+ adam_betas:
100
+ - 0.9
101
+ - 0.95
102
+ max_grad_norm: 1.0
103
+ logging_steps: 10
104
+ schedule:
105
+ max_lr: 0.0001
106
+ lr_init: 0.0
107
+ lr_min: 1.0e-08
108
+ pct_start: 0.1
109
+ anneal: cos
110
+ lr_decay: 0.875
111
+ one_cycle: true
112
+ all_linear: true
113
+ target_modules:
114
+ - q_proj
115
+ - k_proj
116
+ - v_proj
117
+ - o_proj
118
+ include_prefixes:
119
+ - language_model
120
+ - model.language_model
121
+ exclude_prefixes:
122
+ - vision_tower
123
+ - multi_modal_projector
124
+ - vision_model
125
+ vision_adapter:
126
+ enabled: true
127
+ include_prefixes:
128
+ - vision_tower
129
+ - vision_model
130
+ layers: last_third
131
+ all_linear: true
132
+ method: qdora
133
+ rank: 4
134
+ lora_dropout: 0.05
135
+ bias: none
136
+ alpha_mode: proportional
137
+ alpha_ratio: 2.0
138
+ alpha_fixed: 16
139
+ quant:
140
+ load_in_4bit: true
141
+ quant_type: nf4
142
+ double_quant: true
143
+ compute_dtype: bfloat16
144
+ validation:
145
+ every_steps: 40
146
+ patience: 3
147
+ min_delta: 0.005
148
+ exp31:
149
+ E: 4
150
+ c:
151
+ - 1
152
+ - 2
153
+ - a
154
+ baseline: true
155
+ exp32:
156
+ E:
157
+ - 2
158
+ - 4
159
+ - 6
160
+ c: null
161
+ baseline: true
162
+ analysis:
163
+ collapse_threshold: 0.9
164
+ static_cv: 0.05
165
+ metrics:
166
+ normalize:
167
+ lowercase: true
168
+ collapse_whitespace: true
169
+ strip_section_headers: false
170
+ bleu_orders:
171
+ - 1
172
+ - 2
173
+ - 3
174
+ - 4
175
+ rouge_types:
176
+ - rouge1
177
+ - rouge2
178
+ - rougeL
179
+ bertscore:
180
+ enabled: true
181
+ model_type: roberta-large
182
+ num_layers: 17
183
+ batch_size: 32
184
+ rescale_with_baseline: false
185
+ keep:
186
+ - f1
187
+ chexbert:
188
+ enabled: true
189
+ device: auto
190
+ radgraph:
191
+ enabled: true
192
+ model_type: radgraph-xl
193
+ reward_level: all
194
+ batch_size: 16
195
+ device: auto
196
+ primary: chexbert_f1
197
+ selection:
198
+ clinical:
199
+ - chexbert_f1
200
+ - radgraph_f1_entity
201
+ - radgraph_f1_partial
202
+ - radgraph_f1_complete
203
+ text:
204
+ - bleu1
205
+ - bleu2
206
+ - bleu3
207
+ - bleu4
208
+ - rouge1
209
+ - rouge2
210
+ - rougeL
211
+ - bertscore_f1
212
+ tie_tol: 0.0
213
+ stats:
214
+ alpha: 0.05
215
+ n_boot: 4
216
+ bootstrap_sample_size: 6400
217
+ bootstrap_seeds:
218
+ - 42
219
+ - 43
220
+ - 44
221
+ - 45
222
+ ci_method: percentile
223
+ ci: 0.95
224
+ zero_method: pratt
225
+ posthoc_permutation: false
226
+ kendall_w_floor: 0.05
227
+ rrb_floor: 0.1
228
+ wandb:
229
+ enabled: true
230
+ project: rgexp-section3
231
+ entity: null
232
+ run_name: null
233
+ group: null
234
+ resume_same_run: false
235
+ mode: online
236
+ tags:
237
+ - ec-router
238
+ - cls-head
239
+ - moe-dora
240
+ - mimic-cxr
241
+ notes: ''
242
+ log_every: 20
243
+ log_samples: 20
244
+ output:
245
+ dir: ./runs
246
+ hub_repo: RGExp123/Research_results
247
+ hub_repo_type: dataset
248
+ hub_private: true
249
+ push: true
250
+ push_per_arm: true
251
+ subdir: 3.1_c_vs_baseline_v4
252
+ adapter_dir: ./adapters_3_1_v4
3.1_c_vs_baseline_v4/logs/section3_20261105_204504.log ADDED
@@ -0,0 +1,539 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ⚙ Tự kiểm chứng §3 ...
2
+ ✓ config: 12800/128/6400, 1 patient–1 mẫu enforce, split hub v4
3
+ ✓ đúng cỡ 6400/128/12800
4
+ ✓ cân bằng nhãn: train~test ≤ 0.002 · cặp có val (128 mẫu) ≤ 1/128 ở cả 8 seed (cỡ thật) [tệ nhất 0.0048]
5
+ ✓ tốt hơn chia ngẫu nhiên ≥ 5 lần [ngẫu nhiên TB 0.0683 vs 0.0042]
6
+ ✓ hình dạng phân phối: JS hồ sơ nhãn & cardinality ≤ ngưỡng theo cặp (val: 0.06/128)
7
+ ✓ hồ sơ prevalence: Pearson ≥ 0.999 giữa mọi cặp tập
8
+ ✓ balance_report phát hiện split lệch
9
+ ✓ audit: hợp lệ ⇒ ok
10
+ ✓ audit: patient chung giữa hai tập ⇒ không ok
11
+ ✓ audit: patient 2 ảnh trong cùng tập ⇒ không ok
12
+ pool: 1300 patient (mỗi patient 1 ảnh) từ 1500 patient · gán nhãn CheXbert ...
13
+ [nhãn pool] CheXbert gán nhãn 1300/1300 báo cáo (còn lại lấy từ cache pool.npz)
14
+ split train: 800 ảnh / 800 patient
15
+ split val: 100 ảnh / 100 patient
16
+ split test: 400 ảnh / 400 patient
17
+ cân bằng nhãn: max|Δrate| = 0.0075 (ngưỡng 0.02) · ĐẠT
18
+ ✓ build: cỡ 800/100/400 + 1 bệnh nhân – 1 mẫu toàn cục (≈1300 patient, mỗi ảnh một patient)
19
+ ✓ build: balance_report.json + BALANCE.md được ghi, rep ok
20
+ ✓ cache nhãn train/val/test được gieo từ pool ⇒ không gán nhãn lại [gọi thêm 0]
21
+ pool: 1300 patient (mỗi patient 1 ảnh) từ 1500 patient · gán nhãn CheXbert ...
22
+ [nhãn pool] CheXbert gán nhãn 1300/1300 báo cáo (còn lại lấy từ cache pool.npz)
23
+ split train: 800 ảnh / 800 patient
24
+ split val: 100 ảnh / 100 patient
25
+ split test: 400 ảnh / 400 patient
26
+ cân bằng nhãn: max|Δrate| = 0.0075 (ngưỡng 0.02) · ĐẠT
27
+ ✓ tái lập: cùng seed + cùng nhãn ⇒ cùng split
28
+ ✓ thiếu patient (cần 2500 > 1500) ⇒ DỪNG
29
+ pool: 1300 patient (mỗi patient 1 ảnh) từ 1500 patient · gán nhãn CheXbert ...
30
+ split train: 800 ảnh / 800 patient
31
+ split val: 100 ảnh / 100 patient
32
+ split test: 400 ảnh / 400 patient
33
+ cân bằng nhãn: max|Δrate| = 0.4138 (ngưỡng 0.002) · KHÔNG ĐẠT
34
+ ✓ split lệch + enforce ⇒ DỪNG
35
+ ✓ split cũ val 1280 vs cấu hình 128 ⇒ bị đánh dấu CŨ [val: file 1280 ≠ cấu hình 128]
36
+ ✓ split đúng cỡ ⇒ không bị đánh dấu cũ
37
+
38
+ ✓ TẤT CẢ 18/18
39
+ ✓ Friedman χ² khớp scipy (không ties) [46.4757 vs 46.4757]
40
+ ✓ Friedman: arm mạnh nhất có hạng TB tốt nhất (higher_is_better)
41
+ ✓ Friedman: H0 đúng ⇒ không bác (p_F > 0.01), Kendall W ≈ 0
42
+ ✓ Friedman: ties dày vẫn hữu hạn, p ∈ [0,1], có thống kê ties
43
+ ✓ Friedman: mọi khối trùng hoàn toàn ⇒ suy biến, p = 1
44
+ ✓ Wilcoxon p khớp scipy (xấp xỉ chuẩn, hiệu chỉnh liên tục) [0.06191 vs 0.06191]
45
+ ✓ Holm: đơn điệu, ≥ p thô, ≤ 1
46
+ ✓ friedman_all: metric có tín hiệu bác H0, metric nhiễu thuần không bác [{'chexbert_f1': True, 'bleu1': False}]
47
+ ✓ hậu kiểm CÓ CỔNG: chỉ metric đã bác H0, đủ 21 cặp
48
+ ✓ hậu kiểm: baseline (kém hơn 0.12) thua có ý nghĩa & đáng kể ở cả 6 cặp
49
+ ✓ hậu kiểm: không có cột permutation khi posthoc_permutation=false
50
+ ✓ báo cáo: MỘT bảng tóm tắt (hạng TB + ▲▼ + kết luận) và các bảng chi tiết nằm trong <details> thu gọn
51
+ ✓ báo cáo: ngoài khối thu gọn chỉ có đúng một bảng (bảng tóm tắt) [1]
52
+ ✓ báo cáo: metric bác H0 hiện ▲▼, metric không bác không hiện
53
+ ✓ báo cáo: không metric nào bác H0 ⇒ ghi rõ không chạy hậu kiểm
54
+ ✓ báo cáo: thiếu Friedman ⇒ nhắc chạy --stages stats
55
+
56
+ ✓ TẤT CẢ 16/16
57
+ /home/ezycloudx-admin/RGExp_proj/scripts/selftest_ec.py:233: UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
58
+ Consider using tensor.detach() first. (Triggered internally at /pytorch/torch/csrc/autograd/generated/python_variable_methods.cpp:836.)
59
+ refk[b_, e_, t_] = min(math.ceil(max((t_ + 1) / E, math.exp(float(Hh))) - 1e-6), t_ + 1)
60
+ ✓ bước 0: MoE-DoRA == base
61
+ ✓ Σ wᵢ·DoRAᵢ khớp ma trận đầy đủ (bias=False) max|Δ|=3.58e-07
62
+ ✓ Σ wᵢ·DoRAᵢ khớp ma trận đầy đủ (bias=True) max|Δ|=2.53e-07
63
+ ✓ Δw bước 0 = vector E số 0
64
+ ✓ Δw: ‖Δ(A,B,m)‖ từng expert khớp tính tay; expert đứng yên = 0 max|Δ|=0.0e+00
65
+ ✓ Δw: không MoE ⇒ snapshot None, Δw None
66
+ ✓ w one-hot == DoRA expert 1
67
+ ✓ gradient tới lora_A/B/magnitude
68
+ ✓ eval: cache chuẩn, kết quả lặp lại
69
+ ✓ token-wise: Σ w·DoRA + (1−Σw)·W₀ khớp ma trận đầy đủ max|Δ|=2.38e-07
70
+ ✓ token không expert nào chọn == base
71
+ ✓ bước giải mã 1 token == base
72
+ ✓ token-wise: gradient tới A/B/magnitude
73
+ ✓ EC c=1.0: weights [B,T,E]
74
+ ✓ EC c=1.0: token t được expert e chọn ⇔ z_te nằm trong top-⌈c·t/E⌉ của {z_se, s≤t}
75
+ ✓ EC c=1.0: Σ_e w ∈ {0,1}
76
+ ✓ EC c=1.0: token đầu luôn được chọn (k_1 = 1)
77
+ ✓ EC c=1.0: route_w [B,E]
78
+ ✓ EC c=1.5: weights [B,T,E]
79
+ ✓ EC c=1.5: token t được expert e chọn ⇔ z_te nằm trong top-⌈c·t/E⌉ của {z_se, s≤t}
80
+ ✓ EC c=1.5: Σ_e w ∈ {0,1}
81
+ ✓ EC c=1.5: token đầu luôn được chọn (k_1 = 1)
82
+ ✓ EC c=1.5: route_w [B,E]
83
+ ✓ nhân quả: đổi token từ vị trí 30 KHÔNG đổi định tuyến vị trí < 30
84
+ ✓ nhân quả: định tuyến vị trí ≥ 30 có đổi (kiểm chứng test không vô nghĩa)
85
+ ✓ c=1.5 phủ nhiều token hơn c=1
86
+ ✓ EC-Router: gradient tới Norm + Linear
87
+ ✓ pad trái: cho cùng kết quả như chuỗi không pad
88
+ ✓ RouterStream (sinh từng token) == định tuyến teacher-forced toàn chuỗi
89
+ ✓ Cls head [B,T,14,4]
90
+ ✓ Cls head nhân quả: đổi tương lai không đổi logit vị trí < 30
91
+ ✓ cls loss stream hữu hạn, ô toàn 0 bị bỏ, có gradient
92
+ ✓ loss chỉ tính vị trí hợp lệ
93
+ ✓ loss một vị trí [B,L,C] vẫn dùng được
94
+ ✓ lưu/nạp Heads giữ c, E và đầu ra
95
+ ✓ Heads không route ⇒ không Router
96
+ ✓ adaptive: k_ti ≥ sàn c·n_t/E (= k của c=1) ở mọi (t, expert)
97
+ ✓ adaptive: k khớp tính tay ⌈max(n/E, exp H)⌉ (clamp ≤ n)
98
+ ✓ expert đồng đều (chưa chuyên biệt, N̂ = n) ⇒ k = n
99
+ ✓ expert nhọn (đã chuyên biệt, N̂ ≈ 1) ⇒ k về sàn ⌈n/E⌉
100
+ ✓ adaptive nhân quả: đổi tương lai không đổi route quá khứ
101
+ ✓ adaptive: Σ_e w ∈ {0,1}
102
+ ✓ c_eff [B,E]; c cố định ≈ c, adaptive ≥ c cố định [4.0, 4.0, 4.0, 4.0]
103
+ ✓ adaptive: RouterStream (từng token) == teacher-forced
104
+ ✓ lưu/nạp Heads giữ adaptive
105
+
106
+ 45 kiểm tra đạt
107
+ ✓ 3.1: dò c ở E=4 cố định ⇒ c1_e4, c2_e4, ac_e4 + Baseline [['c1_e4', 'c2_e4', 'ac_e4', 'baseline_infer']]
108
+ ✓ 3.1: c số ⇒ capacity cố định k=⌈c·N/E⌉ (c=1,2) E=4; c=a ⇒ Adaptive Capacity E=4
109
+ ✓ 3.1: Baseline = Gemma gốc, kind=infer (không train)
110
+ ✓ 3.2: c tốt nhất của 3.1 = ac_e4 ⇒ c = a [a]
111
+ ✓ 3.2: c tốt nhất của 3.1 = c2_e4 ⇒ c = 2.0 [2.0]
112
+ ✓ 3.2: c tốt nhất của 3.1 = c1_e4 ⇒ c = 1.0 [1.0]
113
+ ✓ 3.2: thiếu điểm của một arm c ở 3.1 ⇒ chưa chọn được (None)
114
+ ✓ 3.2: người dùng ép exp32.c thắng quy tắc tự động (không cần điểm 3.1)
115
+ ✓ 3.2: dò E ∈ {2,4,6} ở c=a ⇒ ac_e2, ac_e4, ac_e6 + Baseline; E4 và Baseline kế thừa từ 3.1
116
+ ✓ 3.2: c=2 ⇒ c2_e2, c2_e4, c2_e6
117
+ ✓ 3.2 không có c ⇒ DỪNG
118
+ ⚙ kế thừa từ 3.1: ac_e4 → đủ predictions + điểm
119
+ ✓ 3.2: inherit_arm kéo predictions + điểm + báo cáo con từ 3.1 và trỏ adapter 3.1
120
+ ✓ 3.3 chưa thiết lập
121
+ ✓ 3.4 chưa thiết lập
122
+ ✓ 3.1-1 chưa thiết lập
123
+ ✓ thư mục kết quả/adapter v4 (Hub không tái dùng kết quả v3 chạy trên data khác)
124
+ ✓ data: train 12800 · val 128 · test 6400, split MỚI (không split_ref/ cũ)
125
+ ✓ bootstrap: 4 tập × 6400 mẫu (lấy có hoàn lại) cho bảng điểm, 4 seed khác nhau
126
+ ✓ kiểm định = Friedman + hậu kiểm Wilcoxon (pratt, không permutation); không còn cấu hình ghép cặp cũ
127
+ ✓ infer greedy
128
+ ✓ cân bằng nhãn bắt buộc (enforce) trong cấu hình split
129
+ ✓ model_cfg: runner gemma3_ec + adapter_dir
130
+ ✓ model_cfg: baseline dùng runner gemma3
131
+ ✓ thang batch infer theo bội 2
132
+ ✓ micro-batch dò theo ước của eff_bs
133
+ ✓ xếp hạng: lâm sàng ngang nhau (a,b) ⇒ văn bản tốt hơn (b) đứng trên; c lâm sàng kém không lên dù văn bản cao
134
+ ✓ xếp hạng: chexbert hơn nhau nhưng 3 metric RadGraph ngang ⇒ hạng lâm sàng TB quyết định
135
+ ✓ xếp hạng: đúng một arm mạnh nhất mọi metric lâm sàng thắng dù văn bản kém
136
+ ✓ lịch finetune: batch 32 · 400 bước (1 lượt qua 12800 mẫu) [(32, 1, False, 400, 0)]
137
+ ✓ không còn pretrain/Mentor trong config
138
+ ✓ val mỗi 40 bước · patience 3
139
+ ✓ lr OneCycle: 0 → 1e-4 → ≈1e-8 [đáy=1.190e-08]
140
+ ✓ OneCycle: một chu kỳ duy nhất, không khởi động lại ở bước 200 [lr[200]=5.869e-05]
141
+ ✓ weight_decay = 0.001 (train)
142
+ ✓ route_corr: expert phân hoá < 0.5 [-0.332]
143
+ ✓ verdict: route tĩnh (đồng đều) -> SỤP ĐỔ [cv=0.0004]
144
+ ✓ verdict: tham số trùng nhau -> SỤP ĐỔ
145
+ ✓ verdict: route động + tham số khác nhau -> KHÔNG sụp đổ [cv=1.16]
146
+ ✓ param_cos: 3 expert giống hệt -> 1
147
+ ✓ param_cos khớp ΔW dựng tường minh (Gram trick) [max|Δ|=1.19e-07]
148
+ ✓ expert_labels: E0 cao nhất trên Cardiomegaly, E1 trên Pleural Effusion
149
+ ✓ expert_labels: bệnh không có ca dương -> NaN
150
+ ✓ E-Clinical: E0 tương quan dương với Cardiomegaly, E1 với Pleural Effusion
151
+ ✓ E-E = C·Cᵀ: đối xứng, đúng công thức
152
+ ✓ md_side_by_side: 2 bảng cạnh nhau
153
+ ✓ heatmap E-Clinical | E-E: ghi PNG
154
+ ✓ md_expert_labels: n/a + in đậm
155
+ [nhãn t] CheXbert gán nhãn 2/2 báo cáo (còn lại lấy từ cache t.npz)
156
+ [nhãn t] CheXbert gán nhãn 1/3 báo cáo (còn lại lấy từ cache t.npz)
157
+ ✓ nhãn lazy: lần 2 chỉ gán key mới
158
+ ✓ onehot (N,14,4) tổng 1
159
+ ✓ positives = pos|unc
160
+ ✓ báo cáo con /tmp/tmpbeik9_a3/run_rp/arms/base/REPORT.md
161
+ ✓ báo cáo con /tmp/tmpbeik9_a3/run_rp/arms/dora/REPORT.md
162
+ ✓ báo cáo con: có arms/<id>/REPORT.md + summary.json
163
+ ✓ báo cáo cuối thiếu báo cáo con ⇒ DỪNG (ReportIncomplete) [{'moe': 'chưa có summary.json'}]
164
+ ✓ báo cáo con /tmp/tmpbeik9_a3/run_rp/arms/moe/REPORT.md
165
+ ✓ arm MoE chưa có route ⇒ báo cáo con CHƯA ĐỦ
166
+ ✓ báo cáo con /tmp/tmpbeik9_a3/run_rp/arms/moe/REPORT.md
167
+ ✓ /tmp/tmpbeik9_a3/run_rp/REPORT.md
168
+ ✓ báo cáo cuối đủ: lấy TRỰC TIẾP từ báo cáo con (giá trị sửa tay xuất hiện) + liên kết tới arms/<id>
169
+
170
+ ✓ TẤT CẢ 54/54
171
+ [VRAM] khởi động §3: trống 15.3 / 15.5 GiB
172
+
173
+ ═══ §3.1 — c1_e4 · c2_e4 · ac_e4 · baseline_infer ═══
174
+ ⚙ Hub: quét RGExp123/Research_results — 206 tệp
175
+ ⚙ trạng thái: c1_e4[infer✓] · c2_e4[infer✓] · ac_e4[infer✓] · baseline_infer[cần train]
176
+ wandb: [wandb.login()] Loaded credentials for https://api.forge.coreweave.com from WANDB_API_KEY.
177
+ wandb: Currently logged in as: ktnminecraft (ktnminecraft-i-h-c-qu-c-gia-tp-hcm) to https://api.forge.coreweave.com. Use `wandb login --relogin` to force relogin
178
+ wandb: setting up run r7fho2po
179
+ wandb: Tracking run with wandb version 0.31.0
180
+ wandb: Run data is saved locally in /home/ezycloudx-admin/RGExp_proj/wandb/run-20261105_204517-r7fho2po
181
+ wandb: Run `wandb offline` to turn off syncing.
182
+ wandb: Syncing run 3.1_c_vs_baseline_v4 · 11-05 20:45
183
+ wandb: ⭐️ View project at https://forge.coreweave.com/wandb/ktnminecraft-i-h-c-qu-c-gia-tp-hcm/rgexp-section3
184
+ wandb: 🚀 View run at https://forge.coreweave.com/wandb/ktnminecraft-i-h-c-qu-c-gia-tp-hcm/rgexp-section3/runs/r7fho2po
185
+ ✓ W&B: https://forge.coreweave.com/wandb/ktnminecraft-i-h-c-qu-c-gia-tp-hcm/rgexp-section3/runs/r7fho2po
186
+ ⚙ dùng lại manifest: 6400 test
187
+
188
+ ═══ Sinh báo cáo (GREEDY) ═══
189
+ ⚙ c1_e4: đã có predictions đủ — bỏ qua infer.
190
+ ⚙ c2_e4: đã có predictions đủ — bỏ qua infer.
191
+ ⚙ ac_e4: đã có predictions đủ — bỏ qua infer.
192
+
193
+ ── Baseline · Gemma 3 gốc (gemma3) ──
194
+ [VRAM] trước infer: trống 15.3 / 15.5 GiB
195
+ resume: đã có 124 mẫu trong baseline_infer.jsonl
196
+ gemma3: còn 6276/6400 mẫu cần sinh
197
+ nạp google/gemma-3-4b-it ...
198
+ Using a slow image processor as `use_fast` is unset and a slow processor was saved with this model. `use_fast=True` will be the default behavior in v4.52, even if the model was saved with a slow processor. This will result in minor differences in outputs. You'll still be able to use a slow processor with `use_fast=False`.
199
+ ảnh: ảnh nguồn <= 512px · processor 896² · pan&scan=tắt -> 256 token/ảnh
200
+
201
+ ✓ google/gemma-3-4b-it
202
+ ⚙ dịch thử mã C OK (/usr/bin/cc) -> để torch.compile chạy
203
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
204
+ ✓ sinh thử OK: Findings: The lungs are clear bilaterally with no evidence of focal consolidation, effusion, or pneumothorax. The heart size is within normal limits. The medias...
205
+ [VRAM] trước infer: trống 7.0 / 15.5 GiB
206
+ dò batch sinh: bắt đầu 512 (hiệu chỉnh theo VRAM trống ở batch đầu)
207
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
208
+ [VRAM] trống 7.0 GiB · ~174 MiB/mẫu → batch tối đa ≈ 24; dò xuống từ đó.
209
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
210
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
211
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
212
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
213
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
214
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
215
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
216
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
217
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
218
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
219
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
220
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
221
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
222
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
223
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
224
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
225
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
226
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
227
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
228
+ [gemma3] 460/6276 | 1.51 ảnh/s | còn ~64.2 phút
229
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
230
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
231
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
232
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
233
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
234
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
235
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
236
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
237
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
238
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
239
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
240
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
241
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
242
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
243
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
244
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
245
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
246
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
247
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
248
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
249
+ [gemma3] 940/6276 | 1.53 ảnh/s | còn ~58.2 phút
250
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
251
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
252
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
253
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
254
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
255
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
256
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
257
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
258
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
259
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
260
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
261
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
262
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
263
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
264
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
265
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
266
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
267
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
268
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
269
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
270
+ [gemma3] 1420/6276 | 1.54 ảnh/s | còn ~52.7 phút
271
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
272
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
273
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
274
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
275
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
276
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
277
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
278
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
279
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
280
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
281
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
282
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
283
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
284
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
285
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
286
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
287
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
288
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
289
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
290
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
291
+ [gemma3] 1900/6276 | 1.54 ảnh/s | còn ~47.3 phút
292
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
293
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
294
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
295
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
296
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
297
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
298
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
299
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
300
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
301
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
302
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
303
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
304
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
305
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
306
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
307
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
308
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
309
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
310
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
311
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
312
+ [gemma3] 2380/6276 | 1.54 ảnh/s | còn ~42.1 phút
313
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
314
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
315
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
316
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
317
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
318
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
319
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
320
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
321
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
322
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
323
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
324
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
325
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
326
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
327
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
328
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
329
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
330
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
331
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
332
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
333
+ [gemma3] 2860/6276 | 1.54 ảnh/s | còn ~36.9 phút
334
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
335
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
336
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
337
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
338
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
339
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
340
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
341
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
342
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
343
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
344
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
345
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
346
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
347
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
348
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
349
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
350
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
351
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
352
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
353
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
354
+ [gemma3] 3340/6276 | 1.55 ảnh/s | còn ~31.6 phút
355
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
356
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
357
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
358
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
359
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
360
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
361
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
362
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
363
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
364
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
365
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
366
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
367
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
368
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
369
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
370
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
371
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
372
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
373
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
374
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
375
+ [gemma3] 3820/6276 | 1.54 ảnh/s | còn ~26.5 phút
376
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
377
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
378
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
379
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
380
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
381
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
382
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
383
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
384
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
385
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
386
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
387
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
388
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
389
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
390
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
391
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
392
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
393
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
394
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
395
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
396
+ [gemma3] 4300/6276 | 1.55 ảnh/s | còn ~21.3 phút
397
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
398
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
399
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
400
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
401
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
402
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
403
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
404
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
405
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
406
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
407
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
408
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
409
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
410
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
411
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
412
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
413
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
414
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
415
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
416
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
417
+ [gemma3] 4780/6276 | 1.55 ảnh/s | còn ~16.1 phút
418
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
419
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
420
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
421
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
422
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
423
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
424
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
425
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
426
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
427
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
428
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
429
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
430
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
431
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
432
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
433
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
434
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
435
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
436
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
437
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
438
+ [gemma3] 5260/6276 | 1.55 ảnh/s | còn ~10.9 phút
439
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
440
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
441
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
442
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
443
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
444
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
445
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
446
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
447
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
448
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
449
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
450
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
451
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
452
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
453
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
454
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
455
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
456
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
457
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
458
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
459
+ [gemma3] 5740/6276 | 1.55 ảnh/s | còn ~5.8 phút
460
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
461
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
462
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
463
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
464
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
465
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
466
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
467
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
468
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
469
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
470
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
471
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
472
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
473
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
474
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
475
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
476
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
477
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
478
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
479
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
480
+ [gemma3] 6220/6276 | 1.55 ảnh/s | còn ~0.6 phút
481
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
482
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
483
+ The following generation flags are not valid and may be ignored: ['top_p', 'top_k']. Set `TRANSFORMERS_VERBOSITY=info` for more details.
484
+ [gemma3] 6276/6276 | 1.55 ảnh/s | còn ~0.0 phút
485
+ ✓ gemma3: xong sau 67.4 phút -> runs/3.1_c_vs_baseline_v4/predictions/baseline_infer.jsonl
486
+ ✓ đã đẩy 3.1_c_vs_baseline_v4/predictions/baseline_infer.jsonl -> RGExp123/Research_results
487
+ [c1_e4] BLEU ...
488
+ [c1_e4] ROUGE ...
489
+ [c1_e4] BERTScore (roberta-large) ...
490
+ Some weights of RobertaModel were not initialized from the model checkpoint at roberta-large and are newly initialized: ['pooler.dense.bias', 'pooler.dense.weight']
491
+ You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
492
+ [c1_e4] CheXbert clinical F1 ...
493
+ [c1_e4] RadGraph F1 (radgraph-xl) ...
494
+ /home/ezycloudx-admin/RGExp_proj/.venv/lib/python3.12/site-packages/huggingface_hub/file_download.py:1727: DeprecationWarning: hf_xet.download_files() is deprecated. Use XetSession().new_file_download_group().start_download_file() instead.
495
+ xet_get(
496
+ /usr/lib/python3.12/tarfile.py:2301: DeprecationWarning: Python 3.14 will, by default, filter extracted tar archives and reject files or modify their metadata. Use the filter argument to control this behavior.
497
+ warnings.warn(
498
+ Using device: cuda:0
499
+ [c2_e4] BLEU ...
500
+ [c2_e4] ROUGE ...
501
+ [c2_e4] BERTScore (roberta-large) ...
502
+ Some weights of RobertaModel were not initialized from the model checkpoint at roberta-large and are newly initialized: ['pooler.dense.bias', 'pooler.dense.weight']
503
+ You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
504
+ [c2_e4] CheXbert clinical F1 ...
505
+ [c2_e4] RadGraph F1 (radgraph-xl) ...
506
+ [ac_e4] BLEU ...
507
+ [ac_e4] ROUGE ...
508
+ [ac_e4] BERTScore (roberta-large) ...
509
+ Some weights of RobertaModel were not initialized from the model checkpoint at roberta-large and are newly initialized: ['pooler.dense.bias', 'pooler.dense.weight']
510
+ You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
511
+ [ac_e4] CheXbert clinical F1 ...
512
+ [ac_e4] RadGraph F1 (radgraph-xl) ...
513
+ [baseline_infer] BLEU ...
514
+ [baseline_infer] ROUGE ...
515
+ [baseline_infer] BERTScore (roberta-large) ...
516
+ Some weights of RobertaModel were not initialized from the model checkpoint at roberta-large and are newly initialized: ['pooler.dense.bias', 'pooler.dense.weight']
517
+ You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
518
+ [baseline_infer] CheXbert clinical F1 ...
519
+ [baseline_infer] RadGraph F1 (radgraph-xl) ...
520
+ ⚙ xếp hạng #1: ac_e4 (hạng lâm sàng TB 1.00, văn bản 1.12) — chỉ là gợi ý theo hạng trung bình (best.json)
521
+
522
+ ═══ Friedman (tổng thể, mọi metric) → hậu kiểm có cổng ═══
523
+ Friedman trên 4 arm × 6400 khối: 12/12 metric bác H0 (Holm trên 12 metric)
524
+ hậu kiểm 72 cặp×metric (chỉ metric đã qua cổng): 59 có ý nghĩa, 56 có ý nghĩa và đáng kể
525
+
526
+ ═══ Phân tích Route: Similarity giữa Expert · Expert-Labels ═══
527
+ [c1_e4] route_corr=-0.139 param_cos=0.073 -> có phân hoá giữa các expert
528
+ [c2_e4] route_corr=-0.266 param_cos=0.108 -> có phân hoá giữa các expert
529
+ [ac_e4] route_corr=-0.277 param_cos=0.109 -> có phân hoá giữa các expert
530
+
531
+ ═══ Báo cáo ═══
532
+ ✓ báo cáo con runs/3.1_c_vs_baseline_v4/arms/c1_e4/REPORT.md
533
+ ✓ báo cáo con runs/3.1_c_vs_baseline_v4/arms/c2_e4/REPORT.md
534
+ ✓ báo cáo con runs/3.1_c_vs_baseline_v4/arms/ac_e4/REPORT.md
535
+ ✓ báo cáo con runs/3.1_c_vs_baseline_v4/arms/baseline_infer/REPORT.md
536
+ ✓ runs/3.1_c_vs_baseline_v4/REPORT.md
537
+ wandb: WARNING Symlinked 1 file into the W&B run directory; call wandb.save again to sync new files.
538
+
539
+ ═══ Đẩy lên Hugging Face ═══
3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_e_clinical_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,0.029001990028797858,0.027284941347483677,0.010297911329383999,0.01007343913995225,0.023650292934116662,0.026346032738128414,0.009358882635251205,0.004257906313219938,0.00224755178824077,0.033779593378869335,-0.01773659597821672,-0.010203219566820632,0.06695860859306262,0.008552539969362772
3
+ E1,0.013149081643942758,0.011316749892327347,-0.002630078240569749,-0.0060997687069583246,0.009321234257296055,-0.025659469048783287,-0.002031506161697511,-0.028878098161536195,-0.001894222704743286,-0.03871303728781846,-0.0005436320380950803,0.01407555604224104,-0.07017143874114694,-0.0017351947548875624
4
+ E2,-0.022981609038170536,-0.014948283062809558,-0.012578696277354069,0.00755789582553853,-0.03758045929393058,-0.018007723624421462,-0.008586793863235813,-0.011409530630742452,-0.020012122251719482,0.0037436990184401622,0.0007210645789809787,-0.002308193404422158,-0.06997667390181819,0.007707783526132676
5
+ E3,0.001494350211941346,-0.006833921743423601,0.005967846165088293,-0.009691384812355036,0.014237149645816086,0.034217474600459284,0.0013650611300677468,0.03511663309766913,0.023921815041397038,0.014029556168968818,0.00011523776591231422,-0.014578293074283013,0.11637725258145656,-0.0013850560003802852
3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_e_e_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,0.009273679000033248,-0.006155315776942644,-0.007147315642929719,0.0096762059761657
3
+ E1,-0.006155315776942644,0.008556202861925887,0.0047314162748929465,-0.010734094390851451
4
+ E2,-0.007147315642929719,0.0047314162748929465,0.00828392318632153,-0.010190944627756894
5
+ E3,0.0096762059761657,-0.010734094390851451,-0.010190944627756894,0.017314251894865813
3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_e_heatmap.png ADDED

Git LFS Details

  • SHA256: 33105d13e45811e310435c30407537065af67d4e09bcce87d93ddaa5e40f946a
  • Pointer size: 131 Bytes
  • Size of remote file: 133 kB
3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_expert_labels_lift.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,1.0043998961263498,1.0033681409176276,1.0012098708494537,1.0028684355525486,1.0043992403286457,1.0071530965619142,1.0014858453569693,1.000480036387081,1.0010541258568337,1.0050271516526752,0.9929991202481365,0.9970631622321255,1.009687930326464,1.0007969358659754
3
+ E1,1.002760942190754,1.0019334645480071,0.9995723329412137,0.9975960359763724,1.0023997270827225,0.9903578520101851,0.9995536094203197,0.9954939717314704,0.9987704079025722,0.9920260761903429,0.9997030149042266,1.005607322708054,0.9859481911265898,0.9997762187203351
4
+ E2,0.9973681542272512,0.9986070872271527,0.9988844452886061,1.0016245519737246,0.9947232243962028,0.9963093483430489,0.9989709273722955,0.9990290179207284,0.9929149766656089,1.0004205657227812,1.000214843335288,0.9994984891899423,0.9923573624948844,1.0005421546012263
5
+ E3,1.0003503660476791,0.9986962562786315,1.001083583752421,0.9957351067248255,1.0040927876596961,1.0143575988150453,1.0003349320522972,1.006118513560256,1.0173393130176196,1.0032267587853374,1.0000702962527326,0.9935150833174867,1.0260224202599466,0.9998005423981698
3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_expert_labels_mean.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,0.09700666702470455,0.0969070183346066,0.09669856890459352,0.0968587559322033,0.09700660368663595,0.09727257585139311,0.09672522299651558,0.09662808023872688,0.09668352678571425,0.09706724844074842,0.09590556050955415,0.09629807266435991,0.09751739444995033,0.09665868693467335
3
+ E1,0.23928608055853917,0.2390886218487393,0.23852519151943455,0.23805359322033898,0.23919988479262658,0.23632636532507745,0.2385207235772357,0.23755198342175035,0.23833383035714292,0.2367244490644489,0.23855637579617825,0.23996530449826992,0.23527410009910787,0.2385738442211054
4
+ E2,0.49131502577873326,0.4919253384262795,0.4920619681978792,0.4934117762711864,0.49001210291858704,0.49079344582043327,0.49210457026713234,0.4921331863395223,0.4891213392857143,0.4928186787941792,0.49271733757961755,0.49236445328719713,0.48884665213082273,0.4928785748743723
5
+ E3,0.16436113748657355,0.16408936134453797,0.16448160777385182,0.163602833898305,0.1649760307219664,0.1666625758513932,0.16435860162601607,0.16530886472148545,0.16715248214285713,0.16483373908523896,0.16431512101910833,0.1632380761245675,0.16857914767096147,0.16427079949748738
3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_route_verdict.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "route_corr_offdiag": -0.2766297042401296,
3
+ "param_cos_offdiag": 0.10896808910684253,
4
+ "route_cv": 0.07330342882834558,
5
+ "threshold": 0.9,
6
+ "cv_floor": 0.05,
7
+ "collapsed_by_param": false,
8
+ "collapsed_by_route": false,
9
+ "collapsed": false,
10
+ "reading": "có phân hoá giữa các expert",
11
+ "n_images": 6400,
12
+ "mean_weight": {
13
+ "E0": 0.0965817175000001,
14
+ "E1": 0.2386272445312495,
15
+ "E2": 0.4926115032812514,
16
+ "E3": 0.16430357109374988
17
+ },
18
+ "top1_share": {
19
+ "E0": 0.0,
20
+ "E1": 0.0,
21
+ "E2": 1.0,
22
+ "E3": 0.0
23
+ }
24
+ }
3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_sim_param_cos.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,1.0,0.09075724318377672,0.09438470378079523,0.11511747932686495
3
+ E1,0.09075724318377672,1.0,0.09305293962410341,0.12644578877931759
4
+ E2,0.09438470378079523,0.09305293962410342,1.0,0.13405037994619715
5
+ E3,0.11511747932686489,0.12644578877931759,0.13405037994619715,1.0
3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_sim_route_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,1.0,-0.18842228526983973,-0.15764880558453082,0.06089919910891525
3
+ E1,-0.18842228526983973,1.0,-0.64630030723273,-0.22770535232204556
4
+ E2,-0.1576488055845308,-0.64630030723273,1.0,-0.5006006741405467
5
+ E3,0.060899199108915246,-0.22770535232204556,-0.5006006741405467,0.9999999999999998
3.1_c_vs_baseline_v4/metrics/analysis/ac_e4_sim_route_cos.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,0.9999999999999986,0.9933063984330308,0.9964696447415624,0.9937757156233795
3
+ E1,0.9933063984330308,1.000000000000001,0.9925253586490355,0.9897469649272391
4
+ E2,0.9964696447415624,0.9925253586490355,0.9999999999999962,0.991969402830465
5
+ E3,0.9937757156233795,0.9897469649272391,0.991969402830465,1.0000000000000022
3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_e_clinical_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,0.020802602671289583,-0.0048122384070191475,-0.016545230157883884,0.024412019598496596,-0.03233482815187352,0.005184547594760089,-0.013824617721318417,-0.03728982823402704,0.006609487291886123,-0.0008012556369530595,-0.013457426744585365,-0.01163733714032374,-0.030492927953746807,0.03066511417163733
3
+ E1,-0.020916341474653172,-0.015132093344330223,0.0009880655125392567,-0.02141402355167992,-0.005207085955001309,-0.03360727902531349,0.011275043893673039,-0.031038008179371582,-0.03175570787506757,-0.05415911557123908,-0.00012395535510693724,0.012586149462917886,-0.1078066856476099,0.038248511667539814
4
+ E2,0.04223480751727722,0.09035062805266311,0.06123167250990046,-0.010578698575576069,0.09558577473535945,0.0315623671552873,0.02956957029452947,0.07907448218083776,0.014366820039710189,0.06740220261706717,-0.010172372284838313,0.010108766371818683,0.09696745783769822,-0.09898284792035915
5
+ E3,0.002045298059343228,-0.0087277243050781,-0.009631795907836832,-0.0011572155690196082,-0.016362141240546378,-0.007443036493606882,0.01003872897362044,0.010934751536764955,0.00397382713898063,-0.00026755228783946326,-0.006018877491987675,0.00938983167566698,-0.013897710359303667,-0.005048718337458104
3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_e_e_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,0.006210694413990362,0.004243199759873831,-0.013044045509177296,0.0004267670354173167
3
+ E1,0.004243199759873831,0.020558349202412764,-0.02386079095159684,0.0015260725927867503
4
+ E2,-0.013044045509177296,-0.02386079095159684,0.05522361692229719,-0.002569794421163723
5
+ E3,0.0004267670354173167,0.0015260725927867503,-0.002569794421163723,0.0010768250063779745
3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_e_heatmap.png ADDED

Git LFS Details

  • SHA256: 9f74e19ced7e26ac26d48a27580899b18e5c0fc1e9bb0a98a1cba90496068a0c
  • Pointer size: 131 Bytes
  • Size of remote file: 137 kB
3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_expert_labels_lift.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,1.0066662472397971,0.998745231863539,0.995894067590772,1.0146831811770194,0.9872954046003031,1.0029733007762323,0.9953639099882782,0.9911199061058352,1.0065478639717156,0.9997481234216143,0.9887799837167103,0.9929246902389891,0.9906809244632352,1.006035621043155
3
+ E1,0.9946445273783002,0.9968474375692931,1.0001959174236885,0.9897088700793514,0.9983653176794147,0.984600391940442,1.0030211014335448,0.9940943275531999,0.9748636617309103,0.986396940800242,0.9999174256632664,1.006114107412214,0.9736750390169914,1.0060150640456207
4
+ E2,1.0149098787116686,1.0259529883143743,1.0167399807269626,0.9929904758071912,1.0413736079375964,1.0199405620222604,1.0109240477643107,1.020744519516384,1.0156795060525314,1.0233416001246372,0.9906568449218951,1.0067706419870728,1.0326467206386323,0.9785376549973821
5
+ E3,1.000560130068031,0.998055150931073,0.9979572520097243,0.9994051612490105,0.9945058696849661,0.9963520685868404,1.0028770406297087,1.0022253823111111,1.0033644136440523,0.9999281223460486,0.9957114003678272,1.0048788580945023,0.9963701765828383,0.9991507666164894
3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_expert_labels_mean.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,0.1218883071965628,0.12092922154316245,0.12058399929328592,0.12285900677966098,0.11954286328725042,0.12144116099071188,0.12051980720092904,0.12000593832891214,0.12187397321428572,0.1210506528066529,0.11972261783439492,0.12022446366782016,0.11995278592666006,0.12181195025125616
3
+ E1,0.1681517572502685,0.16852417494270416,0.16909025936395747,0.16731734915254245,0.1687807834101382,0.1664537244582043,0.1695678768873403,0.16805874204244017,0.16480766071428565,0.16675744386694383,0.1690431783439491,0.1700907716262976,0.16460671555996015,0.1700740276381911
4
+ E2,0.18979133619763697,0.1918564323911384,0.1901335710247347,0.1856923389830508,0.19474013671274967,0.1907320897832818,0.18904597328687564,0.19088243236074248,0.18993525892857147,0.19136809459459428,0.18525594267515924,0.18826927335640142,0.19310818136769078,0.18298961608040185
5
+ E3,0.1339965531686359,0.13366108250572944,0.13364797173144868,0.13384187796610164,0.13318575729646684,0.13343300309597522,0.13430683739837404,0.13421956631299728,0.13437210714285705,0.13391191372141364,0.13334720382165613,0.13457492387543252,0.1334354281466799,0.13380780904522618
3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_route_verdict.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "route_corr_offdiag": -0.13875327534314474,
3
+ "param_cos_offdiag": 0.07319301052722915,
4
+ "route_cv": 0.12412635149631922,
5
+ "threshold": 0.9,
6
+ "cv_floor": 0.05,
7
+ "collapsed_by_param": false,
8
+ "collapsed_by_route": false,
9
+ "collapsed": false,
10
+ "reading": "có phân hoá giữa các expert",
11
+ "n_images": 6400,
12
+ "mean_weight": {
13
+ "E0": 0.12108115031250062,
14
+ "E1": 0.16905713812499987,
15
+ "E2": 0.1870031420312501,
16
+ "E3": 0.13392153968749995
17
+ },
18
+ "top1_share": {
19
+ "E0": 0.00421875,
20
+ "E1": 0.30234375,
21
+ "E2": 0.68234375,
22
+ "E3": 0.01109375
23
+ }
24
+ }
3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_sim_param_cos.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,1.0,0.07981046445669862,0.05178462233814019,0.09684721863923969
3
+ E1,0.07981046445669862,1.0,0.04483095357645377,0.10231830218278833
4
+ E2,0.05178462233814019,0.04483095357645377,1.0,0.06356650197005456
5
+ E3,0.09684721863923966,0.10231830218278833,0.06356650197005459,1.0
3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_sim_route_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,0.9999999999999998,-0.15762665439378168,-0.1301051390138185,-0.10812918207075962
3
+ E1,-0.15762665439378168,1.0,-0.30432903382394283,-0.031243565172664124
4
+ E2,-0.1301051390138185,-0.30432903382394283,0.9999999999999999,-0.10108607758390173
5
+ E3,-0.10812918207075964,-0.03124356517266412,-0.10108607758390173,1.0
3.1_c_vs_baseline_v4/metrics/analysis/c1_e4_sim_route_cos.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,1.0000000000000009,0.9837158161636244,0.9785626900179504,0.9835123366960292
3
+ E1,0.9837158161636244,1.0000000000000018,0.9794740347673,0.9878094350895792
4
+ E2,0.9785626900179504,0.9794740347673,1.0000000000000016,0.9816650692451665
5
+ E3,0.9835123366960292,0.9878094350895792,0.9816650692451665,0.9999999999999991
3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_e_clinical_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,-0.007930492975845953,-0.04881091065676136,-0.03791276661676145,0.004952127864247708,-0.06455421957073118,-0.01382495857331236,-0.03736636727923224,-0.05092123453155257,0.022159097189434967,-0.029115130219275108,0.012191628462464682,-0.012000258177457993,-0.00828926244453073,0.06472117573409376
3
+ E1,-0.015576385135676975,-0.023929937884146435,-0.03351492647449298,-0.00350610730974592,-0.04526702650645237,-0.03964694784142206,-0.014502180767403516,-0.04288372521222152,-0.03399313803968396,-0.07826525900090851,0.0027131046128350154,0.008391644534461541,-0.18220310235068882,0.04084404019741215
4
+ E2,0.049256177805843356,0.10508498675236425,0.0739901963255877,-0.029963804525854735,0.11258389313152836,0.04480839435558625,0.0526032490597194,0.08097728952900618,0.0035921344565115177,0.10609956779355957,-0.013481732872184085,0.013406455311193798,0.13297483343745184,-0.11905355608961328
5
+ E3,-0.0007437161481991803,-0.024406732305714187,-0.007048606567558903,0.00765275786416813,-0.012906889465381516,0.017618747119828606,-0.0037002439072465897,0.017579284538730894,0.017057222349827074,0.012222427059557905,0.008925871365510743,-0.0019275822565562185,0.024445868530364845,0.009911620900084201
3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_e_e_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,0.01814379429181335,0.014352357246028246,-0.03459207163519112,0.0019279850290365465
3
+ E1,0.014352357246028246,0.04984609104381689,-0.05420610252216201,-0.0055870996434518785
4
+ E2,-0.03459207163519112,-0.05420610252216201,0.08733632008171109,0.0004955499734603486
5
+ E3,0.0019279850290365465,-0.0055870996434518785,0.0004955499734603486,0.0027237843762124864
3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_e_heatmap.png ADDED

Git LFS Details

  • SHA256: 7ef79585a6a138a10ab99b0b4ab9b152e4b80764d8e8c1a05e1bbdb7e818f949
  • Pointer size: 131 Bytes
  • Size of remote file: 140 kB
3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_expert_labels_lift.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,0.9974901563220556,0.9874305578804936,0.9907080426921446,1.002941650297948,0.9749505807229244,0.9921697692138379,0.9876245114700019,0.9880240811521153,1.0216803697372263,0.990961046142812,1.0100386642067596,0.9927944825031589,0.9974980862168942,1.0125807496892618
3
+ E1,0.997062058833819,0.9963274256806355,0.9951045803259357,0.9987587641034813,0.9895314887492432,0.9866170944482489,0.9971375051261766,0.9939891989383314,0.9801785322787628,0.985519005489229,1.0013314065143526,1.0030029741662758,0.9672250417830259,1.004731716848008
4
+ E2,1.011874034358522,1.0206125095949685,1.0138129682267005,0.9864422614974933,1.0332767109398495,1.0193313066764274,1.0132704346587966,1.0145065540719183,1.0026770612713178,1.0250902016714492,0.9915442714854302,1.006131682474556,1.0305715184085573,0.9823723832216446
5
+ E3,0.999838019950948,0.9956746967989321,0.9988111342239108,1.0031284195561967,0.9965533082518595,1.0068674194288363,0.9991566270789113,1.0028452402428767,1.0114849936527748,1.0026113477694372,1.0050579281835457,0.9992034824778918,1.0050777316828232,1.0013259056855657
3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_expert_labels_mean.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,Enlarged Cardiomediastinum,Cardiomegaly,Lung Opacity,Lung Lesion,Edema,Consolidation,Pneumonia,Atelectasis,Pneumothorax,Pleural Effusion,Pleural Other,Fracture,Support Devices,No Finding
2
+ E0,0.15876790332975316,0.15716674331550798,0.15768841201413414,0.15963560338983052,0.15518033794162803,0.1579210712074303,0.1571976144018584,0.1572612128647216,0.16261819642857137,0.15772868191268208,0.16076521656050957,0.15802050519031133,0.15876916551040643,0.16116983366834153
3
+ E1,0.28461119441460814,0.28440149350649346,0.2840524325088335,0.2850955186440678,0.28246159447004615,0.2816296811145512,0.2846327305458768,0.28373404708222805,0.27979179464285714,0.281316231808732,0.2858298789808919,0.286307027681661,0.27609422299306235,0.2868004969849246
4
+ E2,0.29061193984962436,0.29312164476699737,0.2911688049469966,0.2833078915254236,0.2967588248847926,0.2927536811145509,0.2910129883855982,0.29136800397878,0.2879705535714286,0.294407645530146,0.2847731974522295,0.28896272664359873,0.29598188899900923,0.2821390155778893
5
+ E3,0.2023718485499463,0.20152917265087858,0.20216400212014132,0.20303784067796604,0.20170700768049146,0.2037946315789474,0.2022339314750289,0.20298052387267945,0.2047292500000001,0.20293318295218296,0.20342838216560505,0.2022434152249134,0.2034323904856295,0.2026730035175884
3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_route_verdict.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "route_corr_offdiag": -0.2659895499372867,
3
+ "param_cos_offdiag": 0.10780160052949572,
4
+ "route_cv": 0.09943060071331047,
5
+ "threshold": 0.9,
6
+ "cv_floor": 0.05,
7
+ "collapsed_by_param": false,
8
+ "collapsed_by_route": false,
9
+ "collapsed": false,
10
+ "reading": "có phân hoá giữa các expert",
11
+ "n_images": 6400,
12
+ "mean_weight": {
13
+ "E0": 0.1591673885937501,
14
+ "E1": 0.28544982921874923,
15
+ "E2": 0.2872016970312495,
16
+ "E3": 0.2024046340625001
17
+ },
18
+ "top1_share": {
19
+ "E0": 0.00015625,
20
+ "E1": 0.48453125,
21
+ "E2": 0.514375,
22
+ "E3": 0.0009375
23
+ }
24
+ }
3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_sim_param_cos.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,1.0,0.08731245817204668,0.09923133135760945,0.11492889648054234
3
+ E1,0.08731245817204668,1.0,0.08970544227526017,0.11719891262298494
4
+ E2,0.09923133135760945,0.08970544227526017,1.0,0.13843256226853082
5
+ E3,0.11492889648054234,0.11719891262298494,0.13843256226853082,1.0
3.1_c_vs_baseline_v4/metrics/analysis/c2_e4_sim_route_corr.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ ,E0,E1,E2,E3
2
+ E0,1.0,-0.18574761108128227,-0.3486689439389931,-0.1208233886965699
3
+ E1,-0.1857476110812823,1.0,-0.47586223816423645,-0.20099353737863243
4
+ E2,-0.3486689439389931,-0.47586223816423645,1.0,-0.2638415803640056
5
+ E3,-0.1208233886965699,-0.2009935373786324,-0.2638415803640056,0.9999999999999998