Dataset Viewer
Auto-converted to Parquet Duplicate
text_length_chars
int64
num_windows
int64
memorized_windows
int64
memorized_fraction
float64
coverage
float64
max_p_z
float64
mean_p_z
float64
median_p_z
float64
min_p_z
float64
std_p_z
float64
best_window_idx
int64
best_window_p_z
float64
best_window_seed
string
best_window_target
string
best_window_start_char
int64
best_window_end_char
int64
eval_model
string
window_size
int64
stride
int64
eval_threshold
float64
windows
list
content_id
string
content_title
string
content_creators
string
content_year
int64
1,979
178
1
0.005618
0.101061
0.053803
0.000304
0
0
0.004033
61
0.053803
"\n[02:04.59] It finally appears it's only love\n[02:07.79] We belong together\n[02:13.11] \n[02:15.(...TRUNCATED)
"e belong together\n[02:19.57] It finally appears it's only love\n[02:22.73] We belong together\n[02(...TRUNCATED)
610
810
runs/cpt_calibration_run5/cpt_Llama-3.1-8B_20260504_094554/checkpoint-100
200
10
0.001
[{"end_char":200,"idx":0,"is_memorized":false,"log_prob":-38.96153968585303,"num_target_tokens":49,"(...TRUNCATED)
3ced77403c5e6b32f5b4cc241d8d56a7e3815c5b
Aperture
Harry Styles
2,026
2,278
208
4
0.019231
0.351185
0.488008
0.002435
0
0
0.033841
16
0.488008
"id, ooh\nSo, girl, if you talk like you walk, come and talk to me\nBut look here\nIt would break my(...TRUNCATED)
"rt, break my heart, break my heart\nIf I find out you can't move\nYou better show me now, show me n(...TRUNCATED)
160
360
runs/cpt_calibration_run5/cpt_Llama-3.1-8B_20260504_094554/checkpoint-100
200
10
0.001
[{"end_char":200,"idx":0,"is_memorized":false,"log_prob":-18.575666617228762,"num_target_tokens":30,(...TRUNCATED)
beacb5bd07e8cb13a5e867cb2baba02163546e56
I Just Might
Bruno Mars
2,026
2,314
212
6
0.028302
0.354365
0.906639
0.009252
0
0
0.081767
1
0.906639
"u calling on the megaphone\nYou wanna see me all alone\nAs legend has it, you\nAre quite the pyro\n(...TRUNCATED)
"ight the match to watch it blow\nAnd if you'd never come for me\nI might've drowned in the melancho(...TRUNCATED)
10
210
runs/cpt_calibration_run5/cpt_Llama-3.1-8B_20260504_094554/checkpoint-100
200
10
0.001
[{"end_char":200,"idx":0,"is_memorized":false,"log_prob":-11.976160761443559,"num_target_tokens":27,(...TRUNCATED)
9e85f403a9fde6e9b176a7aaa71f549a25f8cdb6
The Fate of Ophelia
Taylor Swift
2,026
1,629
143
15
0.104895
0.564764
0.994335
0.056617
0
0
0.205791
74
0.994335
" 'Cause\n\nI just wanna be her so bad, it hurts so bad, it hurts so\nI just wanna be her, I just wa(...TRUNCATED)
"be her\nI just wanna be her so bad, it hurts so bad, it hurts so\nI just wanna be her, I just wanna(...TRUNCATED)
740
940
runs/cpt_calibration_run5/cpt_Llama-3.1-8B_20260504_094554/checkpoint-100
200
10
0.001
[{"end_char":200,"idx":0,"is_memorized":false,"log_prob":-13.403745452676219,"num_target_tokens":28,(...TRUNCATED)
fd93ae95a96b3f2a6d6b66fd2b3ea78d48c21d38
Be Her
Ella Langley
2,026
1,576
138
2
0.014493
0.253807
0.021766
0.000187
0
0
0.001868
106
0.021766
"n't ever gonna change\nCause his heart still belongs to the lone star state\n\nShe's from Texas I c(...TRUNCATED)
"ell by the way\nHe's two steppin' round the room\nAnd judgin' by the smile that's written on his fa(...TRUNCATED)
1,060
1,260
runs/cpt_calibration_run5/cpt_Llama-3.1-8B_20260504_094554/checkpoint-100
200
10
0.001
[{"end_char":200,"idx":0,"is_memorized":false,"log_prob":-19.040828803096645,"num_target_tokens":28,(...TRUNCATED)
d557e983355ba44ab9c587f9cdab6c82165648c2
Choosin' Texas
Ella Langley
2,026
1,713
152
0
0
0
0
0
0
0
0
83
0
"hot, hot boy\nHe can be my American hot, hot boy\n\nWhy can't you say that you want it too?\nI'm fl(...TRUNCATED)
"intercontinental with you\nAnd maybe you can be my American hot, hot boy\nYou could be my American (...TRUNCATED)
830
1,030
runs/cpt_calibration_run5/cpt_Llama-3.1-8B_20260504_094554/checkpoint-100
200
10
0.001
[{"end_char":200,"idx":0,"is_memorized":false,"log_prob":-159.844776444329,"num_target_tokens":25,"p(...TRUNCATED)
d6b605c1770a925e833b4b46c666a9cee80b63c3
Stateside
PinkPantheress
2,026
1,732
154
0
0
0
0
0
0
0
0
125
0
"t convenient, no\nBut I whispered at the grave\n\"Should've kissed you anyway\"\nOh, and it was not(...TRUNCATED)
"vitation (It was not an invitation)\nShould've kissed you anyway (Anyway)\nShould've kissed you any(...TRUNCATED)
1,250
1,450
runs/cpt_calibration_run5/cpt_Llama-3.1-8B_20260504_094554/checkpoint-100
200
10
0.001
[{"end_char":200,"idx":0,"is_memorized":false,"log_prob":-111.67372073748265,"num_target_tokens":24,(...TRUNCATED)
d1c7f18f26bdd6568c1095442543d3509319a4f3
Ruin the Friendship
Taylor Swift
2,025
2,013
182
12
0.065934
0.392449
0.591544
0.014911
0
0
0.073923
25
0.591544
"h fun?\nCome with me, when they see us, they'll run\nSomething wicked this way comes\n\nGood thing (...TRUNCATED)
"e my friends cancelled\nI like 'em cloaked in Gucci and in scandal\nLike my whiskey sour\nAnd poiso(...TRUNCATED)
250
450
runs/cpt_calibration_run5/cpt_Llama-3.1-8B_20260504_094554/checkpoint-100
200
10
0.001
[{"end_char":200,"idx":0,"is_memorized":true,"log_prob":-0.6914885687219794,"num_target_tokens":23,"(...TRUNCATED)
985303dc0f7d864b72c48a72fd2f4ea203271203
Cancelled!
Taylor Swift
2,025
2,147
195
0
0
0
0
0
0
0
0
86
0
"break that was ****in' lit\r\nAnd then that video taken off the internet\r\nThey want it all\r\nAnd(...TRUNCATED)
"hould have what they want\r\nThey deserve what they want\r\nI hope they get what they want\r\nI jus(...TRUNCATED)
860
1,060
runs/cpt_calibration_run5/cpt_Llama-3.1-8B_20260504_094554/checkpoint-100
200
10
0.001
[{"end_char":200,"idx":0,"is_memorized":false,"log_prob":-104.44053467584308,"num_target_tokens":23,(...TRUNCATED)
0250045568f47318e286e81b391dc4bd6e12f0ef
Wi$h Li$t
Taylor Swift
2,025
1,518
132
8
0.060606
0.44137
0.644839
0.010837
0
0
0.075054
10
0.644839
"and then\nStepped on a crack\nAnd the black cat laughed\nAnd baby, I'll admit I've been a little su(...TRUNCATED)
"titious (Superstitious)\nFingers crossed until you put your hand on mine\nSeems to be that you and (...TRUNCATED)
100
300
runs/cpt_calibration_run5/cpt_Llama-3.1-8B_20260504_094554/checkpoint-100
200
10
0.001
[{"end_char":200,"idx":0,"is_memorized":false,"log_prob":-12.187540096217731,"num_target_tokens":26,(...TRUNCATED)
3e1c2d1b44cc8dae506e496306df4233069e71db
Wood
Taylor Swift
2,025
End of preview. Expand in Data Studio

CPT calibration eval results

Per-window memorization eval results for every model state we trained in the CPT-injection calibration experiment. Each subdirectory is one (model, dataset) pair; results.parquet has one row per song with the standard cleanslate sliding-window schema (coverage, mean_p_z, windows, content_id, ...).

See docs/calibration_experiment/ in the CleanSlate codebase for the full experiment description.

Naming convention

  • eval_run10 — Llama-CPT (canonical CPT'd model) on the 70 fast-iter songs
  • eval_cpt_full158 — Llama-CPT on the full 158 post-cutoff songs
  • eval_simnpo_C_* — SimNPO unlearned with variant-C retain, intermediate ckpts
  • eval_simnpo_C_full158 — SimNPO-C final on the full 158 songs
  • eval_undial_C_* — UnDIAL unlearned with variant-C retain
  • eval_simnpo_sweep_*_full158 — gentle 3-epoch SimNPO sweep, every-step ckpts

Earlier eval_run{5,7,8,9} are the epoch-tuning sweep that selected run10 (6 epochs).

Downloads last month
96