Dataset Viewer
Auto-converted to Parquet Duplicate
phase
int64
1
20
lang
stringlengths
2
2
mult_from
int64
128
704
mult_to
int64
128
736
par
int64
101M
579M
bytes
int64
1.21B
6.95B
step
int64
10k
10k
batch
int64
1
4
block
int64
512
512
tokens
int64
5.12M
20.5M
cum
int64
20.5M
164M
exit_ppl
float64
2.29
6.36
1
en
128
128
100,933,632
1,211,147,355
10,000
4
512
20,480,000
20,480,000
2.29
2
es
128
160
126,101,504
1,511,046,575
10,000
4
512
20,480,000
40,960,000
2.42
3
pl
160
192
151,269,376
1,813,044,655
10,000
4
512
20,480,000
61,440,000
3.04
4
fr
192
224
176,437,248
2,115,042,735
10,000
4
512
20,480,000
81,920,000
2.43
5
de
224
256
201,605,120
2,417,040,815
10,000
1
512
5,120,000
87,040,000
2.73
6
cs
256
288
226,772,992
2,719,038,895
10,000
1
512
5,120,000
92,160,000
3.48
7
da
288
320
251,940,864
3,021,036,975
10,000
1
512
5,120,000
97,280,000
2.85
8
pt
320
352
277,108,736
3,323,035,055
10,000
1
512
5,120,000
102,400,000
2.66
9
fi
352
384
302,276,608
3,625,033,135
10,000
1
512
5,120,000
107,520,000
2.98
10
hu
384
416
327,444,480
3,927,031,215
10,000
1
512
5,120,000
112,640,000
2.88
11
bg
416
448
352,612,352
4,229,029,295
10,000
1
512
5,120,000
117,760,000
6.09
12
it
448
480
377,780,224
4,531,027,407
10,000
1
512
5,120,000
122,880,000
2.89
13
et
480
512
402,948,096
4,833,025,499
10,000
1
512
5,120,000
128,000,000
3.26
14
el
512
544
428,115,968
5,135,023,579
10,000
1
512
5,120,000
133,120,000
6.36
15
sk
544
576
453,283,840
5,437,021,659
10,000
1
512
5,120,000
138,240,000
3.58
16
sv
576
608
478,451,712
5,739,019,763
10,000
1
512
5,120,000
143,360,000
3.08
17
ro
608
640
503,619,584
6,041,017,843
10,000
1
512
5,120,000
148,480,000
3.26
18
nl
640
672
528,787,456
6,343,015,923
10,000
1
512
5,120,000
153,600,000
3.07
19
sl
672
704
553,955,328
6,645,014,015
10,000
1
512
5,120,000
158,720,000
3.5
20
lt
704
736
579,123,200
6,947,012,095
10,000
1
512
5,120,000
163,840,000
3.83

BDH Continual-Learning Phase-1 Checkpoints (RA2b)

40 PyTorch checkpoints of the RA2b route-aware growth ladder: a 20-phase continual-learning run over 20 European languages (europarl corpora), one language per phase, route-aware auxiliary loss (alpha 0.9), append-only growth (+32 mult per phase) with step-end frozen-region restore. For each language both the best-val and the last checkpoint are provided.

Contents

  • checkpoints/ — bdh_europarl_ladRA2b-_best.pt and _last.pt, languages: bg cs da de el en es et fi fr hu it lt nl pl pt ro sk sl sv
  • metadata/ra2b_acquisition.{csv,json} — phase schedule: mult range, tokens, parameters, bytes per phase

Loading

These are pipeline-native PyTorch .pt files, not safetensors, and not loadable via transformers AutoModel. The model architecture and loading code live in the bdh-cl repository (pipeline/, bdh.py). Expect to instantiate the model from that source code and load the state dict directly.

Provenance

Produced by the Saga AI Labs BDH continual-learning research line (phase 1: RA2b decay/interference closure, final report and 20x20 serving matrix). Checkpoint family: bdh_europarl_ladRA2b-*. License: Pathway MIT (see LICENSE of the source repository). No claims about downstream performance are made beyond the published phase-1 measurements.

Downloads last month
51