Dataset Viewer
Auto-converted to Parquet Duplicate
eval_id
stringclasses
10 values
step
int64
0
120
split
stringclasses
2 values
protocol
stringclasses
1 value
kv_bits
int32
4
48
n_problems
int32
200
1k
accuracy
float64
0
0.77
format_rate
float64
0
0.78
truncation_rate
float64
0.22
0.78
mean_length
float64
346
800
t_active_at_eval
float64
0
7.29k
t_eval
float64
431
2.8k
trigger
stringclasses
3 values
run_id
stringlengths
15
29
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
451.37206
step0
chk-s1-bf16-b0d7da178fa7
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.675
0.75
0.28
382.46
4,212.876269
534.497329
every25
chk-s1-bf16-b0d7da178fa7
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.68
0.78
0.235
346.265
5,005.151132
528.198807
every25
chk-s1-bf16-b0d7da178fa7
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.64
0.695
0.31
422.135
5,005.151132
563.880915
standalone
chk-s1-bf16-b0d7da178fa7
30|test_id|greedy|kv16
30
test_id
greedy
16
1,000
0.582
0.671
0.347
443.916
5,005.151132
2,795.79712
standalone
chk-s1-bf16-b0d7da178fa7
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
631.475923
step0
chk-s1-k8v4-0742b960789e
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.645
0.715
0.285
416.395
4,374.008601
585.512572
every25
chk-s1-k8v4-0742b960789e
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.67
0.73
0.27
398.955
5,190.645578
529.15775
every25
chk-s1-k8v4-0742b960789e
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.59
0.665
0.34
464.095
5,190.645578
624.110171
standalone
chk-s1-k8v4-0742b960789e
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
532.903588
step0
chk-s1-k8v4-attn-d87d21b77ff6
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.655
0.7
0.305
465.635
4,516.153732
572.507003
every25
chk-s1-k8v4-attn-d87d21b77ff6
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.57
0.65
0.34
491.755
5,520.248964
579.188352
every25
chk-s1-k8v4-attn-d87d21b77ff6
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.52
0.59
0.41
550.885
5,520.248964
691.186655
standalone
chk-s1-k8v4-attn-d87d21b77ff6
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
596.589913
step0
chk-s1-k8v4-tis-9a64b99b15a3
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.715
0.715
0.275
465.895
4,039.536991
628.585127
every25
chk-s1-k8v4-tis-9a64b99b15a3
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.765
0.78
0.22
425.24
5,189.592779
597.857453
every25
chk-s1-k8v4-tis-9a64b99b15a3
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.695
0.73
0.27
455.17
5,189.592779
679.347165
standalone
chk-s1-k8v4-tis-9a64b99b15a3
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
531.654223
step0
chk-s2-bf16-f8841ee10444
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.595
0.64
0.36
468.895
4,275.177785
517.013479
every25
chk-s2-bf16-f8841ee10444
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.625
0.655
0.35
460.84
5,074.031151
486.50139
every25
chk-s2-bf16-f8841ee10444
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.415
0.47
0.535
621.135
5,074.031151
672.986628
standalone
chk-s2-bf16-f8841ee10444
30|test_id|greedy|kv16
30
test_id
greedy
16
1,000
0.535
0.574
0.428
539.712
5,074.031151
2,497.455779
standalone
chk-s2-bf16-f8841ee10444
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
451.145966
step0
chk-s2-k8v4-41bbd1e6ce9e
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.685
0.745
0.255
408.06
3,480.118986
450.200104
every25
chk-s2-k8v4-41bbd1e6ce9e
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.7
0.755
0.245
385.365
4,151.988504
461.267306
every25
chk-s2-k8v4-41bbd1e6ce9e
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.525
0.625
0.37
491.655
4,151.988504
487.318311
standalone
chk-s2-k8v4-41bbd1e6ce9e
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
477.007755
step0
chk-s2-k8v4-attn-4d8b87ff8e41
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.66
0.695
0.305
431.445
4,309.267657
488.643391
every25
chk-s2-k8v4-attn-4d8b87ff8e41
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.665
0.71
0.295
426.175
5,083.08452
489.333575
every25
chk-s2-k8v4-attn-4d8b87ff8e41
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.62
0.695
0.3
422.055
5,083.08452
626.368168
standalone
chk-s2-k8v4-attn-4d8b87ff8e41
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
468.988136
step0
chk-s2-k8v4-tis-f9f30475d917
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.625
0.69
0.31
447.95
3,294.676037
462.750521
every25
chk-s2-k8v4-tis-f9f30475d917
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.66
0.715
0.285
445.515
3,975.720545
456.99655
every25
chk-s2-k8v4-tis-f9f30475d917
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.62
0.67
0.33
460.965
3,975.720545
651.199765
standalone
chk-s2-k8v4-tis-f9f30475d917
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
444.639969
step0
chk-s3-bf16-ae52468a2800
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.675
0.72
0.27
400.96
3,620.266252
447.72036
every25
chk-s3-bf16-ae52468a2800
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.685
0.71
0.29
399.47
4,320.691966
449.338847
every25
chk-s3-bf16-ae52468a2800
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.65
0.74
0.275
399.035
4,320.691966
558.417343
standalone
chk-s3-bf16-ae52468a2800
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
485.374343
step0
chk-s3-k8v4-2ca07b7c95f3
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.655
0.705
0.295
426.565
4,251.413564
521.869445
every25
chk-s3-k8v4-2ca07b7c95f3
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.72
0.745
0.255
400.19
5,089.012476
517.734444
every25
chk-s3-k8v4-2ca07b7c95f3
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.625
0.705
0.29
452.2
5,089.012476
759.387185
standalone
chk-s3-k8v4-2ca07b7c95f3
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
447.809384
step0
chk-s3-k8v4-attn-3fa17f0162f8
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.57
0.64
0.355
462.54
3,755.627311
450.138228
every25
chk-s3-k8v4-attn-3fa17f0162f8
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.62
0.675
0.325
432.775
4,475.726719
446.606733
every25
chk-s3-k8v4-attn-3fa17f0162f8
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.555
0.61
0.395
492.95
4,475.726719
637.902569
standalone
chk-s3-k8v4-attn-3fa17f0162f8
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
442.524296
step0
chk-s3-k8v4-tis-999ed4f562d1
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.7
0.73
0.275
440.015
3,294.972612
511.333264
every25
chk-s3-k8v4-tis-999ed4f562d1
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.735
0.755
0.26
430.565
4,072.080714
536.720164
every25
chk-s3-k8v4-tis-999ed4f562d1
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.725
0.75
0.245
429.455
4,072.080714
552.345297
standalone
chk-s3-k8v4-tis-999ed4f562d1
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
537.068619
step0
chk-s4-bf16-dd82f34d1996
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.565
0.61
0.39
539.92
4,841.367586
595.061959
every25
chk-s4-bf16-dd82f34d1996
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.61
0.645
0.355
498.155
5,837.41746
639.022432
every25
chk-s4-bf16-dd82f34d1996
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.57
0.635
0.365
486.905
5,837.41746
689.087547
standalone
chk-s4-bf16-dd82f34d1996
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
569.771146
step0
chk-s4-k8v4-attn-d73a63918568
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.575
0.65
0.365
489.455
4,763.337484
604.116645
every25
chk-s4-k8v4-attn-d73a63918568
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.605
0.645
0.355
493.345
5,780.958848
633.276002
every25
chk-s4-k8v4-attn-d73a63918568
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.495
0.565
0.445
569.22
5,780.958848
673.768396
standalone
chk-s4-k8v4-attn-d73a63918568
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
638.12388
step0
chk-s4-k8v4-tis-84638df0f184
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.68
0.74
0.26
404.225
4,493.3624
644.067494
every25
chk-s4-k8v4-tis-84638df0f184
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.7
0.715
0.285
426.865
5,462.60009
643.814436
every25
chk-s4-k8v4-tis-84638df0f184
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.66
0.685
0.315
459.44
5,462.60009
680.801895
standalone
chk-s4-k8v4-tis-84638df0f184
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
653.603276
step0
chk-s5-bf16-daa9f6120557
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.575
0.64
0.36
482.43
5,155.468087
647.594682
every25
chk-s5-bf16-daa9f6120557
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.565
0.595
0.405
521.185
6,160.808062
656.108378
every25
chk-s5-bf16-daa9f6120557
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.455
0.53
0.47
572.15
6,160.808062
908.319119
standalone
chk-s5-bf16-daa9f6120557
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
625.545474
step0
chk-s5-k8v4-attn-1d2e1f7bc9ed
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.685
0.725
0.275
427.885
5,944.438908
863.527502
every25
chk-s5-k8v4-attn-1d2e1f7bc9ed
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.695
0.72
0.28
442.25
7,291.764975
879.923609
every25
chk-s5-k8v4-attn-1d2e1f7bc9ed
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.655
0.705
0.3
447.785
7,291.764975
501.416717
standalone
chk-s5-k8v4-attn-1d2e1f7bc9ed
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
1,361.699517
step0
chk-s5-k8v4-tis-e73b5f9d2634
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.67
0.73
0.275
393.175
4,496.124898
480.962366
every25
chk-s5-k8v4-tis-e73b5f9d2634
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.7
0.735
0.265
390.75
5,235.395769
476.49116
every25
chk-s5-k8v4-tis-e73b5f9d2634
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.6
0.64
0.36
480.83
5,235.395769
833.350465
standalone
chk-s5-k8v4-tis-e73b5f9d2634
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
481.140434
step0
chk-s6-bf16-bda68d8099c7
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.575
0.635
0.36
471.24
3,787.338101
466.111595
every25
chk-s6-bf16-bda68d8099c7
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.625
0.675
0.325
451.35
4,525.621644
468.365103
every25
chk-s6-bf16-bda68d8099c7
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.49
0.56
0.44
554.375
4,525.621644
524.368335
standalone
chk-s6-bf16-bda68d8099c7
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
466.402239
step0
chk-s6-k8v4-attn-1f5887a47ca5
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.59
0.67
0.305
442.785
3,775.464004
459.670466
every25
chk-s6-k8v4-attn-1f5887a47ca5
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.655
0.72
0.26
387.525
4,530.901081
468.726323
every25
chk-s6-k8v4-attn-1f5887a47ca5
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.59
0.66
0.33
467.775
4,530.901081
596.029189
standalone
chk-s6-k8v4-attn-1f5887a47ca5
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
454.021301
step0
chk-s6-k8v4-tis-b3dc5f908887
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.63
0.66
0.325
461.36
3,465.719819
457.925693
every25
chk-s6-k8v4-tis-b3dc5f908887
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.655
0.7
0.295
415.495
4,172.703607
514.886236
every25
chk-s6-k8v4-tis-b3dc5f908887
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.545
0.61
0.39
516.135
4,172.703607
540.282597
standalone
chk-s6-k8v4-tis-b3dc5f908887
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
434.884584
step0
chk-s7-bf16-b132cb7bccc8
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.6
0.67
0.325
445.55
3,768.131396
548.868179
every25
chk-s7-bf16-b132cb7bccc8
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.575
0.645
0.355
466.615
4,560.721715
488.437011
every25
chk-s7-bf16-b132cb7bccc8
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.48
0.58
0.4
513.94
4,560.721715
678.877036
standalone
chk-s7-bf16-b132cb7bccc8
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
482.353758
step0
chk-s7-k8v4-attn-9833808f93ec
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.66
0.725
0.275
440.925
4,447.147792
566.965145
every25
chk-s7-k8v4-attn-9833808f93ec
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.66
0.68
0.32
455.77
5,419.410773
560.756101
every25
chk-s7-k8v4-attn-9833808f93ec
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.615
0.65
0.355
495.94
5,419.410773
579.168498
standalone
chk-s7-k8v4-attn-9833808f93ec
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
564.830019
step0
chk-s7-k8v4-tis-ff6bacc89e84
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.485
0.565
0.43
545.4
4,033.88666
580.132749
every25
chk-s7-k8v4-tis-ff6bacc89e84
30|dev200|greedy|kv16
30
dev200
greedy
16
200
0.56
0.62
0.38
503.98
4,939.804511
578.490678
every25
chk-s7-k8v4-tis-ff6bacc89e84
30|dev200|greedy|kv48
30
dev200
greedy
48
200
0.39
0.505
0.51
617.72
4,939.804511
702.163839
standalone
chk-s7-k8v4-tis-ff6bacc89e84
0|dev200|greedy|kv16
0
dev200
greedy
16
200
0.35
0.555
0.405
529.9
0
573.391479
step0
chk-s8-bf16-2fab375e4337
25|dev200|greedy|kv16
25
dev200
greedy
16
200
0.62
0.65
0.35
470.285
4,533.135134
524.816792
every25
chk-s8-bf16-2fab375e4337
End of preview. Expand in Data Studio

KV-Cache Precision GRPO Run Logs (Dress Rehearsal)

Complete logs of the 37 runs behind a study of KV-cache quantization in GRPO post-training of Qwen3-0.6B (4-bit, MLX) on Countdown: the 33 GRPO runs reported in the paper, the frozen BF16 policy used for the gradient diagnostics, and the 3 Dress Rehearsal warm-ups. Every generated token, completion, update cycle and evaluation is here, one Parquet table per log type with a run_id column, so every number in the paper can be recomputed.

Files

File Rows One row per
tokens.parquet 20,687,101 generated rollout token (behaviour log-prob, KV mode, position)
samples.parquet 64,640 rollout completion (text, reward, length, truncation, format)
steps.parquet 1,010 collect/update cycle (phase timings, MLX memory peaks, reward, clip fraction, TIS weights, behaviour log-ratio, gradient norm)
evals.parquet 33,600 evaluated problem (split, KV cache, correct, format, length, answer)
eval_summary.parquet 144 evaluation (accuracy, format rate, truncation rate, mean length)
events.parquet 276 lifecycle or safety event (preflight, preemption, exit)
run_configs.jsonl 37 run: resolved config, environment, warm-up summary
diagnostics/*.json frozen-policy studies: pilot bias table (r), variance tests A7 / A9 / A10, multilevel kill test A6, memory trace A11, log-prob reuse A8, train-cache x serve-cache screen A12, cache-native exactness checks
docs/ decisions.md (dated log of every decision and result), prereg.md (preregistration, amendments A1-A12), dated results

run_id = config name + first 12 hex of the config's SHA-256. In evals, eval_id is step|split|protocol|kv; kv 16 = BF16 cache, 48 = K8V4, 4 = stock 4-bit. Splits: dev200, test_id (1,000, evaluated once).

Run families

Prefix Runs What
chk-s{1..8}-bf16 8 BF16 rollouts (reference)
chk-s{1..3}-k8v4 3 K8V4 rollouts (8-bit keys, 4-bit values)
chk-s{1..8}-k8v4-tis 8 K8V4 + truncated importance sampling
chk-s{1..8}-k8v4-attn 8 K8V4 + attention-driven 8-bit residual buffer
cn-s{1..3}-k8v4 3 K8V4 with a simulated-K8V4 learner
kvp4-s{1,2}-wu, kvn4-s1-wu 3 Dress Rehearsal GRPO through stock 4-bit KV (plain / simulated learner)
wu-kv4-native-s{1,2}, wu-kv4-plain-s1 3 Dress Rehearsal stage-1 warm-ups (simulated / plain cache)
lap-lr3e-5 1 frozen BF16 policy (20 updates) for the diagnostics

Key results

Stock 4-bit serving of BF16-trained policies: 0/64 dev problems. Dress Rehearsal: 0.675 / 0.695 on dev-200 and 0.617 / 0.630 on the untouched 1,000-problem test split through stock 4-bit KV. K8V4 group-gradient variance 1.56x (preregistered, 95% CI [1.11, 2.18], p = 0.014); K8V4 serving penalty for BF16-trained policies 10.9 points. Hardware: Apple M3 MacBook Air 16 GB, about 90 h of laptop compute. Adapters: corbanyax/Dress-Rehearsal-Qwen3-0.6B-4bit-KV-LoRA and corbanyax/KV-Cache-Precision-GRPO-Qwen3-0.6B-LoRA. Licence CC BY 4.0.

Loading

from datasets import load_dataset
steps = load_dataset("corbanyax/KV-Cache-Precision-GRPO-Run-Logs", "steps", split="train").to_pandas()
evals = load_dataset("corbanyax/KV-Cache-Precision-GRPO-Run-Logs", "evals", split="train").to_pandas()
# tokens is 20.7 M rows; stream it or read the Parquet file directly with pyarrow

Project family

Item Hugging Face Kaggle mirror
Dress Rehearsal adapters (this method) corbanyax/Dress-Rehearsal-Qwen3-0.6B-4bit-KV-LoRA model
KV-cache precision study adapters (33 runs) corbanyax/KV-Cache-Precision-GRPO-Qwen3-0.6B-LoRA model
Countdown-3 puzzles with hash-defined splits corbanyax/Countdown-3-Hash-Splits dataset
Full run logs, diagnostics, decision log corbanyax/KV-Cache-Precision-GRPO-Run-Logs dataset

Paper: Beyond a Single Score: A Multi-Objective Evaluation of KV-Cache Quantization in GRPO Training (under double-blind review, AAAI-27 Workshop on Multi-Objective and Multi-Task Optimization for Robust AI). All items are private until the review period ends.

Downloads last month
51

Collection including corbanyax/KV-Cache-Precision-GRPO-Run-Logs