Datasets:

augustinian-babylm
/

token-embeddings

token_id int64 0 100k	n_rows int64 0 132k	seeded bool 2 classes
0	0	false
1	0	false
2	0	false
3	0	false
4	0	false
5	114	true
6	62	true
7	0	false
8	0	false
9	2	true
10	16	true
11	244	true
12	15	true
13	83	true
14	2	true
15	0	false
16	8,140	true
17	7,497	true
18	52,210	true
19	242	true
20	5	true
21	421	true
22	422	true
23	101	true
24	66	true
25	18	true
26	14	true
27	21	true
28	25	true
29	31	true
30	732	true
31	71	true
32	8	true
33	0	false
34	33	true
35	24	true
36	0	false
37	111	true
38	49	true
39	98	true
40	116	true
41	430	true
42	144	true
43	178	true
44	167	true
45	274	true
46	34	true
47	128	true
48	40	true
49	38	true
50	63	true
51	93	true
52	127	true
53	7	true
54	87	true
55	194	true
56	216	true
57	151	true
58	37	true
59	156	true
60	25	true
61	381	true
62	32	true
63	9	true
64	55	true
65	23	true
66	0	false
67	4	true
68	12	true
69	214	true
70	143	true
71	173	true
72	140	true
73	153	true
74	145	true
75	112	true
76	102	true
77	360	true
78	55	true
79	178	true
80	129	true
81	73	true
82	139	true
83	244	true
84	257	true
85	28	true
86	154	true
87	1,534	true
88	199	true
89	50	true
90	138	true
91	102	true
92	57	true
93	362	true
94	42	true
95	0	false
96	1	true
97	2	true
98	0	false
99	0	false

End of preview. Expand in Data Studio

token-embeddings

Per-token visual embedding tables for the Augustinian BabyLM project: [V, 768] float32 matrices used to initialize the input embedding matrix of a DeBERTa-v3-base masked LM before text training.

Organized as <encoder>/<vocab>/, for encoder in dinov3 / sam / ibot and vocab in 50k / 75k / 100k. Each directory holds E_init.safetensors (the table) and a seeded_mask marking which rows carry visual information, roughly 24-38% of rows depending on vocabulary size.

Unseeded rows are zeros and must be overwritten with the model's own random initialization at load time, not used as-is.

Built by averaging the region features in augustinian-babylm/region-embeddings over every region a word labels, then mean-centering, L2-normalizing, and scaling to the model's initializer standard deviation. The published tables use all-subword attribution (--no-seed_last_subword); rebuild with the same flag for comparability.

Part of https://github.com/bylinina/augustinian_babylm. Paper: https://openreview.net/forum?id=B4TD4XdlwF.

Citation

@inproceedings{bylinina2026augustinian,
  title     = {Augustinian BabyLM: What Ostensive Definition Can and Cannot
               Teach a Small Language Model},
  author    = {Bylinina, Lisa},
  booktitle = {Proceedings of the BabyLM Workshop},
  year      = {2026},
  url       = {https://openreview.net/forum?id=B4TD4XdlwF}
}

Downloads last month: 23