wassname/isokl_steering_calibration
1
0
Fork 0
mirror of https://github.com/wassname/isokl_steering_calibration.git synced 2026-06-27 16:45:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

wip: dense alpha sweep + auto-render figs (Qwen3.5-0.8B w=512); queued OLMo-2/Gemma 4B/Gemma 12B/OLMo-3 7B at w=4096

Browse Source
This commit is contained in:
wassname
2026-05-06 05:37:33 +08:00
parent 0bd7a11d2d
commit bd34b7580c
144 changed files with 342825 additions and 210 deletions
Download Patch File Download Diff File Expand all files Collapse all files
figs/fig_Qwen05_w512_spagetti.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 558 KiB

figs/figs_partial/figure1.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 104 KiB

figs/figs_partial/table.csv
+19
View File
@@ -0,0 +1,19 @@
model_short,method,window,c_mean,c_std,n_seeds,c_cv
Llama-3.2-1B-Instruct,directional_ablation,20,5.119636661337771,0.0,3,0.0
Llama-3.2-1B-Instruct,directional_ablation,50,4.6095092483978455,0.0,3,0.0
Llama-3.2-1B-Instruct,mean_diff,20,3.771688942723309,0.0,3,0.0
Llama-3.2-1B-Instruct,mean_diff,50,3.861664231693977,0.0,3,0.0
Llama-3.2-1B-Instruct,pca,20,3.818178678361504,0.0,3,0.0
Llama-3.2-1B-Instruct,pca,50,3.600290823741986,0.0,3,0.0
Qwen2.5-0.5B-Instruct,directional_ablation,20,7.507819866975713,0.0,3,0.0
Qwen2.5-0.5B-Instruct,directional_ablation,50,7.055664779577401,0.0,3,0.0
Qwen2.5-0.5B-Instruct,mean_diff,20,7.588130746747839,0.0,3,0.0
Qwen2.5-0.5B-Instruct,mean_diff,50,7.561536121211781,0.0,3,0.0
Qwen2.5-0.5B-Instruct,pca,20,8.655517019086593,0.0,3,0.0
Qwen2.5-0.5B-Instruct,pca,50,8.606126777874907,0.0,3,0.0
Qwen3-4B-Instruct-2507,directional_ablation,20,25.600000000000005,0.0,3,0.0
Qwen3-4B-Instruct-2507,directional_ablation,50,22.895205490202148,0.0,3,0.0
Qwen3-4B-Instruct-2507,mean_diff,20,25.600000000000005,0.0,3,0.0
Qwen3-4B-Instruct-2507,mean_diff,50,25.600000000000005,0.0,3,0.0
Qwen3-4B-Instruct-2507,pca,20,23.302283905419525,0.0,3,0.0
Qwen3-4B-Instruct-2507,pca,50,17.99025750455262,0.0,3,0.0
1 model_short method window c_mean c_std n_seeds c_cv
2 Llama-3.2-1B-Instruct directional_ablation 20 5.119636661337771 0.0 3 0.0
3 Llama-3.2-1B-Instruct directional_ablation 50 4.6095092483978455 0.0 3 0.0
4 Llama-3.2-1B-Instruct mean_diff 20 3.771688942723309 0.0 3 0.0
5 Llama-3.2-1B-Instruct mean_diff 50 3.861664231693977 0.0 3 0.0
6 Llama-3.2-1B-Instruct pca 20 3.818178678361504 0.0 3 0.0
7 Llama-3.2-1B-Instruct pca 50 3.600290823741986 0.0 3 0.0
8 Qwen2.5-0.5B-Instruct directional_ablation 20 7.507819866975713 0.0 3 0.0
9 Qwen2.5-0.5B-Instruct directional_ablation 50 7.055664779577401 0.0 3 0.0
10 Qwen2.5-0.5B-Instruct mean_diff 20 7.588130746747839 0.0 3 0.0
11 Qwen2.5-0.5B-Instruct mean_diff 50 7.561536121211781 0.0 3 0.0
12 Qwen2.5-0.5B-Instruct pca 20 8.655517019086593 0.0 3 0.0
13 Qwen2.5-0.5B-Instruct pca 50 8.606126777874907 0.0 3 0.0
14 Qwen3-4B-Instruct-2507 directional_ablation 20 25.600000000000005 0.0 3 0.0
15 Qwen3-4B-Instruct-2507 directional_ablation 50 22.895205490202148 0.0 3 0.0
16 Qwen3-4B-Instruct-2507 mean_diff 20 25.600000000000005 0.0 3 0.0
17 Qwen3-4B-Instruct-2507 mean_diff 50 25.600000000000005 0.0 3 0.0
18 Qwen3-4B-Instruct-2507 pca 20 23.302283905419525 0.0 3 0.0
19 Qwen3-4B-Instruct-2507 pca 50 17.99025750455262 0.0 3 0.0
figs/figs_partial/table.md
+20
View File
@@ -0,0 +1,20 @@
| model_short | method | window | c_mean | c_std | n_seeds | c_cv |
|:-----------------------|:---------------------|---------:|---------:|--------:|----------:|-------:|
| Llama-3.2-1B-Instruct | directional_ablation | 20 | 5.120 | 0.000 | 3 | 0.000 |
| Llama-3.2-1B-Instruct | directional_ablation | 50 | 4.610 | 0.000 | 3 | 0.000 |
| Llama-3.2-1B-Instruct | mean_diff | 20 | 3.772 | 0.000 | 3 | 0.000 |
| Llama-3.2-1B-Instruct | mean_diff | 50 | 3.862 | 0.000 | 3 | 0.000 |
| Llama-3.2-1B-Instruct | pca | 20 | 3.818 | 0.000 | 3 | 0.000 |
| Llama-3.2-1B-Instruct | pca | 50 | 3.600 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | directional_ablation | 20 | 7.508 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | directional_ablation | 50 | 7.056 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | mean_diff | 20 | 7.588 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | mean_diff | 50 | 7.562 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | pca | 20 | 8.656 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | pca | 50 | 8.606 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | directional_ablation | 20 | 25.600 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | directional_ablation | 50 | 22.895 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | mean_diff | 20 | 25.600 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | mean_diff | 50 | 25.600 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | pca | 20 | 23.302 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | pca | 50 | 17.990 | 0.000 | 3 | 0.000 |
figs/figs_qwen05_kl/figure1_kl_only.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 75 KiB

figs/figs_qwen05_kl/table.csv
+19
View File
@@ -0,0 +1,19 @@
model_short,method,window,c_mean,c_std,n_seeds,c_cv
Llama-3.2-1B-Instruct,directional_ablation,20,5.119636661337771,0.0,3,0.0
Llama-3.2-1B-Instruct,directional_ablation,50,4.6095092483978455,0.0,3,0.0
Llama-3.2-1B-Instruct,mean_diff,20,3.771688942723309,0.0,3,0.0
Llama-3.2-1B-Instruct,mean_diff,50,3.861664231693977,0.0,3,0.0
Llama-3.2-1B-Instruct,pca,20,3.818178678361504,0.0,3,0.0
Llama-3.2-1B-Instruct,pca,50,3.600290823741986,0.0,3,0.0
Qwen2.5-0.5B-Instruct,directional_ablation,20,7.507819866975713,0.0,3,0.0
Qwen2.5-0.5B-Instruct,directional_ablation,50,7.055664779577401,0.0,3,0.0
Qwen2.5-0.5B-Instruct,mean_diff,20,7.588130746747839,0.0,3,0.0
Qwen2.5-0.5B-Instruct,mean_diff,50,7.561536121211781,0.0,3,0.0
Qwen2.5-0.5B-Instruct,pca,20,8.655517019086593,0.0,3,0.0
Qwen2.5-0.5B-Instruct,pca,50,8.606126777874907,0.0,3,0.0
Qwen3-4B-Instruct-2507,directional_ablation,20,25.600000000000005,0.0,3,0.0
Qwen3-4B-Instruct-2507,directional_ablation,50,22.895205490202148,0.0,3,0.0
Qwen3-4B-Instruct-2507,mean_diff,20,25.600000000000005,0.0,3,0.0
Qwen3-4B-Instruct-2507,mean_diff,50,25.600000000000005,0.0,3,0.0
Qwen3-4B-Instruct-2507,pca,20,23.302283905419525,0.0,3,0.0
Qwen3-4B-Instruct-2507,pca,50,17.99025750455262,0.0,3,0.0
1 model_short method window c_mean c_std n_seeds c_cv
2 Llama-3.2-1B-Instruct directional_ablation 20 5.119636661337771 0.0 3 0.0
3 Llama-3.2-1B-Instruct directional_ablation 50 4.6095092483978455 0.0 3 0.0
4 Llama-3.2-1B-Instruct mean_diff 20 3.771688942723309 0.0 3 0.0
5 Llama-3.2-1B-Instruct mean_diff 50 3.861664231693977 0.0 3 0.0
6 Llama-3.2-1B-Instruct pca 20 3.818178678361504 0.0 3 0.0
7 Llama-3.2-1B-Instruct pca 50 3.600290823741986 0.0 3 0.0
8 Qwen2.5-0.5B-Instruct directional_ablation 20 7.507819866975713 0.0 3 0.0
9 Qwen2.5-0.5B-Instruct directional_ablation 50 7.055664779577401 0.0 3 0.0
10 Qwen2.5-0.5B-Instruct mean_diff 20 7.588130746747839 0.0 3 0.0
11 Qwen2.5-0.5B-Instruct mean_diff 50 7.561536121211781 0.0 3 0.0
12 Qwen2.5-0.5B-Instruct pca 20 8.655517019086593 0.0 3 0.0
13 Qwen2.5-0.5B-Instruct pca 50 8.606126777874907 0.0 3 0.0
14 Qwen3-4B-Instruct-2507 directional_ablation 20 25.600000000000005 0.0 3 0.0
15 Qwen3-4B-Instruct-2507 directional_ablation 50 22.895205490202148 0.0 3 0.0
16 Qwen3-4B-Instruct-2507 mean_diff 20 25.600000000000005 0.0 3 0.0
17 Qwen3-4B-Instruct-2507 mean_diff 50 25.600000000000005 0.0 3 0.0
18 Qwen3-4B-Instruct-2507 pca 20 23.302283905419525 0.0 3 0.0
19 Qwen3-4B-Instruct-2507 pca 50 17.99025750455262 0.0 3 0.0
figs/figs_qwen05_kl/table.md
+20
View File
@@ -0,0 +1,20 @@
| model_short | method | window | c_mean | c_std | n_seeds | c_cv |
|:-----------------------|:---------------------|---------:|---------:|--------:|----------:|-------:|
| Llama-3.2-1B-Instruct | directional_ablation | 20 | 5.120 | 0.000 | 3 | 0.000 |
| Llama-3.2-1B-Instruct | directional_ablation | 50 | 4.610 | 0.000 | 3 | 0.000 |
| Llama-3.2-1B-Instruct | mean_diff | 20 | 3.772 | 0.000 | 3 | 0.000 |
| Llama-3.2-1B-Instruct | mean_diff | 50 | 3.862 | 0.000 | 3 | 0.000 |
| Llama-3.2-1B-Instruct | pca | 20 | 3.818 | 0.000 | 3 | 0.000 |
| Llama-3.2-1B-Instruct | pca | 50 | 3.600 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | directional_ablation | 20 | 7.508 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | directional_ablation | 50 | 7.056 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | mean_diff | 20 | 7.588 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | mean_diff | 50 | 7.562 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | pca | 20 | 8.656 | 0.000 | 3 | 0.000 |
| Qwen2.5-0.5B-Instruct | pca | 50 | 8.606 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | directional_ablation | 20 | 25.600 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | directional_ablation | 50 | 22.895 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | mean_diff | 20 | 25.600 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | mean_diff | 50 | 25.600 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | pca | 20 | 23.302 | 0.000 | 3 | 0.000 |
| Qwen3-4B-Instruct-2507 | pca | 50 | 17.990 | 0.000 | 3 | 0.000 |
figs/figs_qwen05_kl_long/figure1_kl_only.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 74 KiB

figs/figs_qwen05_kl_long/table.csv
+4
View File
@@ -0,0 +1,4 @@
model_short,method,window,c_mean,c_std,n_seeds,c_cv
Qwen2.5-0.5B-Instruct,directional_ablation,50,7.0556647795774,,1,
Qwen2.5-0.5B-Instruct,mean_diff,50,7.561536121211781,,1,
Qwen2.5-0.5B-Instruct,pca,50,8.606126777874907,,1,
1 model_short method window c_mean c_std n_seeds c_cv
2 Qwen2.5-0.5B-Instruct directional_ablation 50 7.0556647795774 1
3 Qwen2.5-0.5B-Instruct mean_diff 50 7.561536121211781 1
4 Qwen2.5-0.5B-Instruct pca 50 8.606126777874907 1
figs/figs_qwen05_kl_long/table.md
+5
View File
@@ -0,0 +1,5 @@
| model_short | method | window | c_mean | c_std | n_seeds | c_cv |
|:----------------------|:---------------------|---------:|---------:|:--------|----------:|:-------|
| Qwen2.5-0.5B-Instruct | directional_ablation | 50 | 7.056 | | 1 | |
| Qwen2.5-0.5B-Instruct | mean_diff | 50 | 7.562 | | 1 | |
| Qwen2.5-0.5B-Instruct | pca | 50 | 8.606 | | 1 | |
figs/figs_qwen05_pmass_fix2/figure1.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 117 KiB

figs/figs_qwen05_pmass_fix2/table.csv
+4
View File
@@ -0,0 +1,4 @@
model_short,method,window,c_mean,c_std,n_seeds,c_cv
Qwen2.5-0.5B-Instruct,directional_ablation,50,7.0556647795774,,1,
Qwen2.5-0.5B-Instruct,mean_diff,50,7.561536121211781,,1,
Qwen2.5-0.5B-Instruct,pca,50,8.606126777874907,,1,
1 model_short method window c_mean c_std n_seeds c_cv
2 Qwen2.5-0.5B-Instruct directional_ablation 50 7.0556647795774 1
3 Qwen2.5-0.5B-Instruct mean_diff 50 7.561536121211781 1
4 Qwen2.5-0.5B-Instruct pca 50 8.606126777874907 1
figs/figs_qwen05_pmass_fix2/table.md
+5
View File
@@ -0,0 +1,5 @@
| model_short | method | window | c_mean | c_std | n_seeds | c_cv |
|:----------------------|:---------------------|---------:|---------:|:--------|----------:|:-------|
| Qwen2.5-0.5B-Instruct | directional_ablation | 50 | 7.056 | | 1 | |
| Qwen2.5-0.5B-Instruct | mean_diff | 50 | 7.562 | | 1 | |
| Qwen2.5-0.5B-Instruct | pca | 50 | 8.606 | | 1 | |
figs/figs_qwen05_pretty/figure1.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 360 KiB

figs/figs_qwen05_pretty/table.csv
+4
View File
@@ -0,0 +1,4 @@
model_short,method,window,c_mean,c_std,n_seeds,c_cv
Qwen2.5-0.5B-Instruct,directional_ablation,50,7.0556647795774,,1,
Qwen2.5-0.5B-Instruct,mean_diff,50,7.561536121211781,,1,
Qwen2.5-0.5B-Instruct,pca,50,8.606126777874907,,1,
1 model_short method window c_mean c_std n_seeds c_cv
2 Qwen2.5-0.5B-Instruct directional_ablation 50 7.0556647795774 1
3 Qwen2.5-0.5B-Instruct mean_diff 50 7.561536121211781 1
4 Qwen2.5-0.5B-Instruct pca 50 8.606126777874907 1
figs/figs_qwen05_pretty/table.md
+5
View File
@@ -0,0 +1,5 @@
| model_short | method | window | c_mean | c_std | n_seeds | c_cv |
|:----------------------|:---------------------|---------:|---------:|:--------|----------:|:-------|
| Qwen2.5-0.5B-Instruct | directional_ablation | 50 | 7.056 | | 1 | |
| Qwen2.5-0.5B-Instruct | mean_diff | 50 | 7.562 | | 1 | |
| Qwen2.5-0.5B-Instruct | pca | 50 | 8.606 | | 1 | |
figs/figs_qwen05_pretty_raw/figure1.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 530 KiB

figs/figs_qwen05_pretty_raw/table.csv
+4
View File
@@ -0,0 +1,4 @@
model_short,method,window,c_mean,c_std,n_seeds,c_cv
Qwen2.5-0.5B-Instruct,directional_ablation,50,7.0556647795774,,1,
Qwen2.5-0.5B-Instruct,mean_diff,50,7.561536121211781,,1,
Qwen2.5-0.5B-Instruct,pca,50,8.606126777874907,,1,
1 model_short method window c_mean c_std n_seeds c_cv
2 Qwen2.5-0.5B-Instruct directional_ablation 50 7.0556647795774 1
3 Qwen2.5-0.5B-Instruct mean_diff 50 7.561536121211781 1
4 Qwen2.5-0.5B-Instruct pca 50 8.606126777874907 1
figs/figs_qwen05_pretty_raw/table.md
+5
View File
@@ -0,0 +1,5 @@
| model_short | method | window | c_mean | c_std | n_seeds | c_cv |
|:----------------------|:---------------------|---------:|---------:|:--------|----------:|:-------|
| Qwen2.5-0.5B-Instruct | directional_ablation | 50 | 7.056 | | 1 | |
| Qwen2.5-0.5B-Instruct | mean_diff | 50 | 7.562 | | 1 | |
| Qwen2.5-0.5B-Instruct | pca | 50 | 8.606 | | 1 | |
figs/figs_qwen05_spaghetti/figure1.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 541 KiB

figs/figs_qwen05_spaghetti/table.csv
+4
View File
@@ -0,0 +1,4 @@
model_short,method,window,c_mean,c_std,n_seeds,c_cv
Qwen2.5-0.5B-Instruct,directional_ablation,50,7.0556647795774,,1,
Qwen2.5-0.5B-Instruct,mean_diff,50,7.561536121211781,,1,
Qwen2.5-0.5B-Instruct,pca,50,8.606126777874907,,1,
1 model_short method window c_mean c_std n_seeds c_cv
2 Qwen2.5-0.5B-Instruct directional_ablation 50 7.0556647795774 1
3 Qwen2.5-0.5B-Instruct mean_diff 50 7.561536121211781 1
4 Qwen2.5-0.5B-Instruct pca 50 8.606126777874907 1
figs/figs_qwen05_spaghetti/table.md
+5
View File
@@ -0,0 +1,5 @@
| model_short | method | window | c_mean | c_std | n_seeds | c_cv |
|:----------------------|:---------------------|---------:|---------:|:--------|----------:|:-------|
| Qwen2.5-0.5B-Instruct | directional_ablation | 50 | 7.056 | | 1 | |
| Qwen2.5-0.5B-Instruct | mean_diff | 50 | 7.562 | | 1 | |
| Qwen2.5-0.5B-Instruct | pca | 50 | 8.606 | | 1 | |
figs/figs_qwen05_spaghetti2/figure1.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 615 KiB

figs/figs_qwen05_spaghetti2/table.csv
+4
View File
@@ -0,0 +1,4 @@
model_short,method,window,c_mean,c_std,n_seeds,c_cv
Qwen2.5-0.5B-Instruct,directional_ablation,50,7.0556647795774,,1,
Qwen2.5-0.5B-Instruct,mean_diff,50,7.561536121211781,,1,
Qwen2.5-0.5B-Instruct,pca,50,8.606126777874907,,1,
1 model_short method window c_mean c_std n_seeds c_cv
2 Qwen2.5-0.5B-Instruct directional_ablation 50 7.0556647795774 1
3 Qwen2.5-0.5B-Instruct mean_diff 50 7.561536121211781 1
4 Qwen2.5-0.5B-Instruct pca 50 8.606126777874907 1
figs/figs_qwen05_spaghetti2/table.md
+5
View File
@@ -0,0 +1,5 @@
| model_short | method | window | c_mean | c_std | n_seeds | c_cv |
|:----------------------|:---------------------|---------:|---------:|:--------|----------:|:-------|
| Qwen2.5-0.5B-Instruct | directional_ablation | 50 | 7.056 | | 1 | |
| Qwen2.5-0.5B-Instruct | mean_diff | 50 | 7.562 | | 1 | |
| Qwen2.5-0.5B-Instruct | pca | 50 | 8.606 | | 1 | |
figs/figs_qwen05_survival/survival.md
+14
View File
@@ -0,0 +1,14 @@
| threshold | alpha | n | S(t=20) | S(t=49) | median_death_t |
|------------:|--------:|----:|----------:|----------:|-----------------:|
| 0.500 | 0.500 | 24 | 0.875 | 0.792 | |
| 0.500 | 1.000 | 24 | 0.000 | 0.000 | 0 |
| 0.500 | 2.000 | 24 | 0.000 | 0.000 | 0 |
| 0.500 | 4.000 | 24 | 0.000 | 0.000 | 0 |
| 1.000 | 0.500 | 24 | 1.000 | 1.000 | |
| 1.000 | 1.000 | 24 | 0.500 | 0.333 | 20 |
| 1.000 | 2.000 | 24 | 0.000 | 0.000 | 0 |
| 1.000 | 4.000 | 24 | 0.000 | 0.000 | 0 |
| 2.000 | 0.500 | 24 | 1.000 | 1.000 | |
| 2.000 | 1.000 | 24 | 0.917 | 0.792 | |
| 2.000 | 2.000 | 24 | 0.000 | 0.000 | 0 |
| 2.000 | 4.000 | 24 | 0.000 | 0.000 | 0 |
figs/figs_qwen05_survival/survival.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 77 KiB

figs/figs_qwen05_survival/survival_pmass.md
+14
View File
@@ -0,0 +1,14 @@
| metric | threshold | alpha | n | S_mid | S_end | t_S<=0.5 |
|:---------|------------:|--------:|----:|--------:|--------:|-----------:|
| pmass | 0.500 | 0.500 | 24 | 0.833 | 0.708 | |
| pmass | 0.500 | 1.000 | 24 | 0.917 | 0.875 | |
| pmass | 0.500 | 2.000 | 24 | 1.000 | 0.958 | |
| pmass | 0.500 | 4.000 | 24 | 0.333 | 0.333 | 0 |
| pmass | 0.800 | 0.500 | 24 | 0.667 | 0.583 | |
| pmass | 0.800 | 1.000 | 24 | 0.583 | 0.542 | |
| pmass | 0.800 | 2.000 | 24 | 0.375 | 0.375 | 0 |
| pmass | 0.800 | 4.000 | 24 | 0.333 | 0.292 | 0 |
| pmass | 0.950 | 0.500 | 24 | 0.333 | 0.208 | 5 |
| pmass | 0.950 | 1.000 | 24 | 0.250 | 0.208 | 0 |
| pmass | 0.950 | 2.000 | 24 | 0.333 | 0.292 | 0 |
| pmass | 0.950 | 4.000 | 24 | 0.000 | 0.000 | 0 |
figs/figs_qwen05_survival/survival_pmass.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 73 KiB

figs/figs_qwen05_w50_survival/survival_pmass.md
+11
View File
@@ -0,0 +1,11 @@
| metric | threshold | alpha | n | S_mid | S_end | t_S<=0.5 |
|:---------|------------:|--------:|----:|--------:|--------:|-----------:|
| pmass | 0.500 | 1.000 | 72 | 0.125 | 0.125 | 0 |
| pmass | 0.500 | 2.000 | 72 | 0.000 | 0.000 | 0 |
| pmass | 0.500 | 4.000 | 72 | 0.000 | 0.000 | 0 |
| pmass | 0.800 | 1.000 | 72 | 0.042 | 0.042 | 0 |
| pmass | 0.800 | 2.000 | 72 | 0.000 | 0.000 | 0 |
| pmass | 0.800 | 4.000 | 72 | 0.000 | 0.000 | 0 |
| pmass | 0.950 | 1.000 | 72 | 0.000 | 0.000 | 0 |
| pmass | 0.950 | 2.000 | 72 | 0.000 | 0.000 | 0 |
| pmass | 0.950 | 4.000 | 72 | 0.000 | 0.000 | 0 |
figs/figs_qwen05_w50_survival/survival_pmass.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 65 KiB

figs/figs_qwen05_w512_color/figure1.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 396 KiB

figs/figs_qwen05_w512_color/table.csv
+4
View File
@@ -0,0 +1,4 @@
model_short,method,window,c_mean,c_std,n_seeds,c_cv
Qwen2.5-0.5B-Instruct,directional_ablation,512,8.59831195323767,,1,
Qwen2.5-0.5B-Instruct,mean_diff,512,8.030432886640742,,1,
Qwen2.5-0.5B-Instruct,pca,512,8.232617543390845,,1,
1 model_short method window c_mean c_std n_seeds c_cv
2 Qwen2.5-0.5B-Instruct directional_ablation 512 8.59831195323767 1
3 Qwen2.5-0.5B-Instruct mean_diff 512 8.030432886640742 1
4 Qwen2.5-0.5B-Instruct pca 512 8.232617543390845 1
figs/figs_qwen05_w512_color/table.md
+5
View File
@@ -0,0 +1,5 @@
| model_short | method | window | c_mean | c_std | n_seeds | c_cv |
|:----------------------|:---------------------|---------:|---------:|:--------|----------:|:-------|
| Qwen2.5-0.5B-Instruct | directional_ablation | 512 | 8.598 | | 1 | |
| Qwen2.5-0.5B-Instruct | mean_diff | 512 | 8.030 | | 1 | |
| Qwen2.5-0.5B-Instruct | pca | 512 | 8.233 | | 1 | |
figs/figs_qwen05_w512_kl_alive/kl_alive_spaghetti.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 207 KiB

figs/figs_qwen05_w512_kl_alive/kl_alive_summary.md
+6
View File
@@ -0,0 +1,6 @@
| alpha | n | n_died | n_censored |
|--------:|----:|---------:|-------------:|
| 0.5 | 24 | 24 | 0 |
| 1 | 24 | 24 | 0 |
| 2 | 24 | 24 | 0 |
| 4 | 24 | 23 | 0 |
figs/figs_qwen05_w512_survival/survival_kl.md
+14
View File
@@ -0,0 +1,14 @@
| metric | threshold | alpha | n | S_mid | S_end | t_S<=0.5 |
|:---------|------------:|--------:|----:|--------:|--------:|-----------:|
| kl | 1.000 | 0.500 | 24 | 0.875 | 0.833 | |
| kl | 1.000 | 1.000 | 24 | 0.000 | 0.000 | 8 |
| kl | 1.000 | 2.000 | 24 | 0.000 | 0.000 | 0 |
| kl | 1.000 | 4.000 | 24 | 0.000 | 0.000 | 0 |
| kl | 2.000 | 0.500 | 24 | 1.000 | 1.000 | |
| kl | 2.000 | 1.000 | 24 | 0.250 | 0.208 | 75 |
| kl | 2.000 | 2.000 | 24 | 0.000 | 0.000 | 0 |
| kl | 2.000 | 4.000 | 24 | 0.000 | 0.000 | 0 |
| kl | 4.000 | 0.500 | 24 | 1.000 | 1.000 | |
| kl | 4.000 | 1.000 | 24 | 0.750 | 0.750 | |
| kl | 4.000 | 2.000 | 24 | 0.083 | 0.083 | 2 |
| kl | 4.000 | 4.000 | 24 | 0.000 | 0.000 | 0 |
figs/figs_qwen05_w512_survival/survival_kl.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 71 KiB

figs/figs_qwen05_w512_survival/survival_pmass_eval.md
+14
View File
@@ -0,0 +1,14 @@
| metric | threshold | alpha | n | S_mid | S_end | t_S<=0.5 |
|:-----------|------------:|--------:|----:|--------:|--------:|-----------:|
| pmass_eval | 0.500 | 0.500 | 24 | 0.250 | 0.208 | 3 |
| pmass_eval | 0.500 | 1.000 | 24 | 0.208 | 0.167 | 0 |
| pmass_eval | 0.500 | 2.000 | 24 | 0.292 | 0.250 | 0 |
| pmass_eval | 0.500 | 4.000 | 24 | 0.292 | 0.292 | 0 |
| pmass_eval | 0.800 | 0.500 | 24 | 0.000 | 0.000 | 0 |
| pmass_eval | 0.800 | 1.000 | 24 | 0.000 | 0.000 | 0 |
| pmass_eval | 0.800 | 2.000 | 24 | 0.000 | 0.000 | 0 |
| pmass_eval | 0.800 | 4.000 | 24 | 0.083 | 0.042 | 0 |
| pmass_eval | 0.950 | 0.500 | 24 | 0.000 | 0.000 | 0 |
| pmass_eval | 0.950 | 1.000 | 24 | 0.000 | 0.000 | 0 |
| pmass_eval | 0.950 | 2.000 | 24 | 0.000 | 0.000 | 0 |
| pmass_eval | 0.950 | 4.000 | 24 | 0.042 | 0.000 | 0 |
figs/figs_qwen05_w512_survival/survival_pmass_eval.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 74 KiB

figs/figs_qwen05_w512_survival_v2/survival_pmass_eval.md
+14
View File
@@ -0,0 +1,14 @@
| metric | threshold | alpha | n | S_mid | S_end | t_S<=0.5 |
|:-----------|------------:|--------:|----:|--------:|--------:|-----------:|
| pmass_eval | 0.500 | 0.500 | 24 | 0.438 | 0.419 | 8 |
| pmass_eval | 0.500 | 1.000 | 24 | 0.333 | 0.319 | 0 |
| pmass_eval | 0.500 | 2.000 | 24 | 0.319 | 0.306 | 0 |
| pmass_eval | 0.500 | 4.000 | 24 | 0.319 | 0.319 | 0 |
| pmass_eval | 0.800 | 0.500 | 24 | 0.077 | 0.077 | 0 |
| pmass_eval | 0.800 | 1.000 | 24 | 0.110 | 0.110 | 0 |
| pmass_eval | 0.800 | 2.000 | 24 | 0.212 | 0.212 | 0 |
| pmass_eval | 0.800 | 4.000 | 24 | 0.255 | 0.244 | 0 |
| pmass_eval | 0.950 | 0.500 | 24 | 0.000 | 0.000 | 0 |
| pmass_eval | 0.950 | 1.000 | 24 | 0.000 | 0.000 | 0 |
| pmass_eval | 0.950 | 2.000 | 24 | 0.040 | 0.040 | 0 |
| pmass_eval | 0.950 | 4.000 | 24 | 0.201 | 0.193 | 0 |
figs/figs_qwen05_w512_survival_v2/survival_pmass_eval.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 82 KiB

figs/image.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 59 KiB

figs/kl_alive_spaghetti.png
BIN
View File
Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 281 KiB

figs/table.csv
+37
View File
@@ -0,0 +1,37 @@
model,method,window,alpha,c_star_mean,n_seeds,kl_p95_mean,pmass_mean
Llama-3.2-1B-Instruct,directional_ablation,20,1.0,5.119636661337771,3,1.1908359713852406,NaN
Llama-3.2-1B-Instruct,directional_ablation,20,2.0,5.119636661337771,3,5.850264692306519,0.000022955508256927715
Llama-3.2-1B-Instruct,directional_ablation,50,1.0,4.6095092483978455,3,1.291592478454113,NaN
Llama-3.2-1B-Instruct,directional_ablation,50,2.0,4.6095092483978455,3,5.11867133140564,0.00004614551000320882
Llama-3.2-1B-Instruct,mean_diff,20,1.0,3.771688942723309,3,0.9256407611072064,NaN
Llama-3.2-1B-Instruct,mean_diff,20,2.0,3.771688942723309,3,5.391779696941375,0.00006245920594665222
Llama-3.2-1B-Instruct,mean_diff,50,1.0,3.861664231693977,3,1.3414980980753899,NaN
Llama-3.2-1B-Instruct,mean_diff,50,2.0,3.861664231693977,3,4.656074690818786,0.00007268900522335157
Llama-3.2-1B-Instruct,pca,20,1.0,3.818178678361504,3,0.9329803831875324,NaN
Llama-3.2-1B-Instruct,pca,20,2.0,3.818178678361504,3,5.620841109752655,0.00026052194389194485
Llama-3.2-1B-Instruct,pca,50,1.0,3.600290823741986,3,0.9301029246300458,NaN
Llama-3.2-1B-Instruct,pca,50,2.0,3.600290823741986,3,4.0647015488147735,NaN
Qwen2.5-0.5B-Instruct,directional_ablation,20,1.0,7.507819866975713,3,0.6075146049261093,0.00005063821326984907
Qwen2.5-0.5B-Instruct,directional_ablation,20,2.0,7.507819866975713,3,3.166464865207672,0.00018519468194426735
Qwen2.5-0.5B-Instruct,directional_ablation,50,1.0,7.055664779577401,3,0.5607812261581421,0.000034678878002771604
Qwen2.5-0.5B-Instruct,directional_ablation,50,2.0,7.055664779577401,3,2.2286340260505675,0.00013751945268516218
Qwen2.5-0.5B-Instruct,mean_diff,20,1.0,7.588130746747839,3,0.7048790633678437,0.000040864707125365383
Qwen2.5-0.5B-Instruct,mean_diff,20,2.0,7.588130746747839,3,3.335330218076706,0.00019464152283035218
Qwen2.5-0.5B-Instruct,mean_diff,50,1.0,7.561536121211781,3,0.7623215705156327,0.00003399881875959016
Qwen2.5-0.5B-Instruct,mean_diff,50,2.0,7.561536121211781,3,2.8257888650894163,0.00015105884966198408
Qwen2.5-0.5B-Instruct,pca,20,1.0,8.655517019086593,3,0.9307092409580946,3.041228809275154e-6
Qwen2.5-0.5B-Instruct,pca,20,2.0,8.655517019086593,3,3.7524219751358032,1.6446400348257838e-7
Qwen2.5-0.5B-Instruct,pca,50,1.0,8.606126777874907,3,0.8437561804056167,NaN
Qwen2.5-0.5B-Instruct,pca,50,2.0,8.606126777874907,3,3.1483269047737124,1.4466886552347544e-7
Qwen3-4B-Instruct-2507,directional_ablation,20,1.0,25.600000000000005,3,1.3869682106771506,9.215271860500782e-6
Qwen3-4B-Instruct-2507,directional_ablation,20,2.0,25.600000000000005,3,7.038443911075592,0.00001822325955023185
Qwen3-4B-Instruct-2507,directional_ablation,50,1.0,22.895205490202148,3,1.0069563373062511,NaN
Qwen3-4B-Instruct-2507,directional_ablation,50,2.0,22.895205490202148,3,4.545377564039081,NaN
Qwen3-4B-Instruct-2507,mean_diff,20,1.0,25.600000000000005,3,0.9436350018950179,4.663814388822024e-6
Qwen3-4B-Instruct-2507,mean_diff,20,2.0,25.600000000000005,3,6.434498374164105,0.000016732647531197965
Qwen3-4B-Instruct-2507,mean_diff,50,1.0,25.600000000000005,3,0.9753538948745699,NaN
Qwen3-4B-Instruct-2507,mean_diff,50,2.0,25.600000000000005,3,5.002368605150841,NaN
Qwen3-4B-Instruct-2507,pca,20,1.0,23.302283905419525,3,1.490876998582462,0.00001391104007908428
Qwen3-4B-Instruct-2507,pca,20,2.0,23.302283905419525,3,5.252012262865901,0.000011667380206471789
Qwen3-4B-Instruct-2507,pca,50,1.0,17.99025750455262,3,0.9170716370666468,6.987194066781776e-6
Qwen3-4B-Instruct-2507,pca,50,2.0,17.99025750455262,3,3.4862812616676093,0.00002544457582068353
1 model method window alpha c_star_mean n_seeds kl_p95_mean pmass_mean
2 Llama-3.2-1B-Instruct directional_ablation 20 1.0 5.119636661337771 3 1.1908359713852406 NaN
3 Llama-3.2-1B-Instruct directional_ablation 20 2.0 5.119636661337771 3 5.850264692306519 0.000022955508256927715
4 Llama-3.2-1B-Instruct directional_ablation 50 1.0 4.6095092483978455 3 1.291592478454113 NaN
5 Llama-3.2-1B-Instruct directional_ablation 50 2.0 4.6095092483978455 3 5.11867133140564 0.00004614551000320882
6 Llama-3.2-1B-Instruct mean_diff 20 1.0 3.771688942723309 3 0.9256407611072064 NaN
7 Llama-3.2-1B-Instruct mean_diff 20 2.0 3.771688942723309 3 5.391779696941375 0.00006245920594665222
8 Llama-3.2-1B-Instruct mean_diff 50 1.0 3.861664231693977 3 1.3414980980753899 NaN
9 Llama-3.2-1B-Instruct mean_diff 50 2.0 3.861664231693977 3 4.656074690818786 0.00007268900522335157
10 Llama-3.2-1B-Instruct pca 20 1.0 3.818178678361504 3 0.9329803831875324 NaN
11 Llama-3.2-1B-Instruct pca 20 2.0 3.818178678361504 3 5.620841109752655 0.00026052194389194485
12 Llama-3.2-1B-Instruct pca 50 1.0 3.600290823741986 3 0.9301029246300458 NaN
13 Llama-3.2-1B-Instruct pca 50 2.0 3.600290823741986 3 4.0647015488147735 NaN
14 Qwen2.5-0.5B-Instruct directional_ablation 20 1.0 7.507819866975713 3 0.6075146049261093 0.00005063821326984907
15 Qwen2.5-0.5B-Instruct directional_ablation 20 2.0 7.507819866975713 3 3.166464865207672 0.00018519468194426735
16 Qwen2.5-0.5B-Instruct directional_ablation 50 1.0 7.055664779577401 3 0.5607812261581421 0.000034678878002771604
17 Qwen2.5-0.5B-Instruct directional_ablation 50 2.0 7.055664779577401 3 2.2286340260505675 0.00013751945268516218
18 Qwen2.5-0.5B-Instruct mean_diff 20 1.0 7.588130746747839 3 0.7048790633678437 0.000040864707125365383
19 Qwen2.5-0.5B-Instruct mean_diff 20 2.0 7.588130746747839 3 3.335330218076706 0.00019464152283035218
20 Qwen2.5-0.5B-Instruct mean_diff 50 1.0 7.561536121211781 3 0.7623215705156327 0.00003399881875959016
21 Qwen2.5-0.5B-Instruct mean_diff 50 2.0 7.561536121211781 3 2.8257888650894163 0.00015105884966198408
22 Qwen2.5-0.5B-Instruct pca 20 1.0 8.655517019086593 3 0.9307092409580946 3.041228809275154e-6
23 Qwen2.5-0.5B-Instruct pca 20 2.0 8.655517019086593 3 3.7524219751358032 1.6446400348257838e-7
24 Qwen2.5-0.5B-Instruct pca 50 1.0 8.606126777874907 3 0.8437561804056167 NaN
25 Qwen2.5-0.5B-Instruct pca 50 2.0 8.606126777874907 3 3.1483269047737124 1.4466886552347544e-7
26 Qwen3-4B-Instruct-2507 directional_ablation 20 1.0 25.600000000000005 3 1.3869682106771506 9.215271860500782e-6
27 Qwen3-4B-Instruct-2507 directional_ablation 20 2.0 25.600000000000005 3 7.038443911075592 0.00001822325955023185
28 Qwen3-4B-Instruct-2507 directional_ablation 50 1.0 22.895205490202148 3 1.0069563373062511 NaN
29 Qwen3-4B-Instruct-2507 directional_ablation 50 2.0 22.895205490202148 3 4.545377564039081 NaN
30 Qwen3-4B-Instruct-2507 mean_diff 20 1.0 25.600000000000005 3 0.9436350018950179 4.663814388822024e-6
31 Qwen3-4B-Instruct-2507 mean_diff 20 2.0 25.600000000000005 3 6.434498374164105 0.000016732647531197965
32 Qwen3-4B-Instruct-2507 mean_diff 50 1.0 25.600000000000005 3 0.9753538948745699 NaN
33 Qwen3-4B-Instruct-2507 mean_diff 50 2.0 25.600000000000005 3 5.002368605150841 NaN
34 Qwen3-4B-Instruct-2507 pca 20 1.0 23.302283905419525 3 1.490876998582462 0.00001391104007908428
35 Qwen3-4B-Instruct-2507 pca 20 2.0 23.302283905419525 3 5.252012262865901 0.000011667380206471789
36 Qwen3-4B-Instruct-2507 pca 50 1.0 17.99025750455262 3 0.9170716370666468 6.987194066781776e-6
37 Qwen3-4B-Instruct-2507 pca 50 2.0 17.99025750455262 3 3.4862812616676093 0.00002544457582068353
figs/table.md
+38
View File
@@ -0,0 +1,38 @@
| model | method | window | alpha | c_star_mean | n_seeds | kl_p95_mean | pmass_mean |
|:-----------------------|:---------------------|---------:|--------:|--------------:|----------:|--------------:|-------------:|
| Llama-3.2-1B-Instruct | directional_ablation | 20 | 1.000 | 5.120 | 3 | 1.191 | nan |
| Llama-3.2-1B-Instruct | directional_ablation | 20 | 2.000 | 5.120 | 3 | 5.850 | 0.000 |
| Llama-3.2-1B-Instruct | directional_ablation | 50 | 1.000 | 4.610 | 3 | 1.292 | nan |
| Llama-3.2-1B-Instruct | directional_ablation | 50 | 2.000 | 4.610 | 3 | 5.119 | 0.000 |
| Llama-3.2-1B-Instruct | mean_diff | 20 | 1.000 | 3.772 | 3 | 0.926 | nan |
| Llama-3.2-1B-Instruct | mean_diff | 20 | 2.000 | 3.772 | 3 | 5.392 | 0.000 |
| Llama-3.2-1B-Instruct | mean_diff | 50 | 1.000 | 3.862 | 3 | 1.341 | nan |
| Llama-3.2-1B-Instruct | mean_diff | 50 | 2.000 | 3.862 | 3 | 4.656 | 0.000 |
| Llama-3.2-1B-Instruct | pca | 20 | 1.000 | 3.818 | 3 | 0.933 | nan |
| Llama-3.2-1B-Instruct | pca | 20 | 2.000 | 3.818 | 3 | 5.621 | 0.000 |
| Llama-3.2-1B-Instruct | pca | 50 | 1.000 | 3.600 | 3 | 0.930 | nan |
| Llama-3.2-1B-Instruct | pca | 50 | 2.000 | 3.600 | 3 | 4.065 | nan |
| Qwen2.5-0.5B-Instruct | directional_ablation | 20 | 1.000 | 7.508 | 3 | 0.608 | 0.000 |
| Qwen2.5-0.5B-Instruct | directional_ablation | 20 | 2.000 | 7.508 | 3 | 3.166 | 0.000 |
| Qwen2.5-0.5B-Instruct | directional_ablation | 50 | 1.000 | 7.056 | 3 | 0.561 | 0.000 |
| Qwen2.5-0.5B-Instruct | directional_ablation | 50 | 2.000 | 7.056 | 3 | 2.229 | 0.000 |
| Qwen2.5-0.5B-Instruct | mean_diff | 20 | 1.000 | 7.588 | 3 | 0.705 | 0.000 |
| Qwen2.5-0.5B-Instruct | mean_diff | 20 | 2.000 | 7.588 | 3 | 3.335 | 0.000 |
| Qwen2.5-0.5B-Instruct | mean_diff | 50 | 1.000 | 7.562 | 3 | 0.762 | 0.000 |
| Qwen2.5-0.5B-Instruct | mean_diff | 50 | 2.000 | 7.562 | 3 | 2.826 | 0.000 |
| Qwen2.5-0.5B-Instruct | pca | 20 | 1.000 | 8.656 | 3 | 0.931 | 0.000 |
| Qwen2.5-0.5B-Instruct | pca | 20 | 2.000 | 8.656 | 3 | 3.752 | 0.000 |
| Qwen2.5-0.5B-Instruct | pca | 50 | 1.000 | 8.606 | 3 | 0.844 | nan |
| Qwen2.5-0.5B-Instruct | pca | 50 | 2.000 | 8.606 | 3 | 3.148 | 0.000 |
| Qwen3-4B-Instruct-2507 | directional_ablation | 20 | 1.000 | 25.600 | 3 | 1.387 | 0.000 |
| Qwen3-4B-Instruct-2507 | directional_ablation | 20 | 2.000 | 25.600 | 3 | 7.038 | 0.000 |
| Qwen3-4B-Instruct-2507 | directional_ablation | 50 | 1.000 | 22.895 | 3 | 1.007 | nan |
| Qwen3-4B-Instruct-2507 | directional_ablation | 50 | 2.000 | 22.895 | 3 | 4.545 | nan |
| Qwen3-4B-Instruct-2507 | mean_diff | 20 | 1.000 | 25.600 | 3 | 0.944 | 0.000 |
| Qwen3-4B-Instruct-2507 | mean_diff | 20 | 2.000 | 25.600 | 3 | 6.434 | 0.000 |
| Qwen3-4B-Instruct-2507 | mean_diff | 50 | 1.000 | 25.600 | 3 | 0.975 | nan |
| Qwen3-4B-Instruct-2507 | mean_diff | 50 | 2.000 | 25.600 | 3 | 5.002 | nan |
| Qwen3-4B-Instruct-2507 | pca | 20 | 1.000 | 23.302 | 3 | 1.491 | 0.000 |
| Qwen3-4B-Instruct-2507 | pca | 20 | 2.000 | 23.302 | 3 | 5.252 | 0.000 |
| Qwen3-4B-Instruct-2507 | pca | 50 | 1.000 | 17.990 | 3 | 0.917 | 0.000 |
| Qwen3-4B-Instruct-2507 | pca | 50 | 2.000 | 17.990 | 3 | 3.486 | 0.000 |