Models / roberta

roberta-base

Reconstructed from its own config.json with no weights read. 9.7M downloads on Hugging Face.

Our count against the checkpoint

The left number comes from the graph. The right one is the number of scalars in the published weight files. Nothing on this page was tuned to make them agree.

Derived from structure

124M

124,034,304 parameters

In the published checkpoint

125M

124,697,947 scalars · safetensors.total, read 2026-09-06

Delta

-0.53%

What it costs to run

Cost is a roofline estimate on the priced GPU for 10 epochs at batch 32 over 50,000 samples (assumed; no dataset attached). GPU fit is fp32 weights plus gradients plus two Adam moments (16 bytes per parameter) with 1.3x headroom; activations are not included and grow with batch size.

Layers

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$1.77

Card	Memory
T4 (16GB)	weights + activations	fits
A100 (40GB)	weights + activations	fits
H100 (80GB)	weights + activations	fits

Structure

52 nodes. Output shapes are propagated from the input shape, batch dimension excluded.

	Layer	Type	Output shape
1	Input	Input	1 × 514
2	Embedding	Embedding	1 × 514 × 768
3	Positional_Embedding	Learned Pos Embed	1 × 514 × 768
4	Attention_1	Multi-Head Attention	1 × 514 × 768
5	Add_1	Add	1 × 514 × 768
6	LayerNorm_1_1	LayerNorm	1 × 514 × 768
7	FFN_1	Feed Forward	1 × 514 × 768
8	Attention_2	Multi-Head Attention	1 × 514 × 768
9	Add_2	Add	1 × 514 × 768
10	LayerNorm_2_1	LayerNorm	1 × 514 × 768
11	FFN_2	Feed Forward	1 × 514 × 768
12	Attention_3	Multi-Head Attention	1 × 514 × 768
13	Add_3	Add	1 × 514 × 768
14	LayerNorm_3_1	LayerNorm	1 × 514 × 768
15	FFN_3	Feed Forward	1 × 514 × 768
16	Attention_4	Multi-Head Attention	1 × 514 × 768
17	Add_4	Add	1 × 514 × 768
18	LayerNorm_4_1	LayerNorm	1 × 514 × 768
19	FFN_4	Feed Forward	1 × 514 × 768
20	Attention_5	Multi-Head Attention	1 × 514 × 768
21	Add_5	Add	1 × 514 × 768
22	LayerNorm_5_1	LayerNorm	1 × 514 × 768
23	FFN_5	Feed Forward	1 × 514 × 768
24	Attention_6	Multi-Head Attention	1 × 514 × 768
25	Add_6	Add	1 × 514 × 768
26	LayerNorm_6_1	LayerNorm	1 × 514 × 768
27	FFN_6	Feed Forward	1 × 514 × 768
28	Attention_7	Multi-Head Attention	1 × 514 × 768
29	Add_7	Add	1 × 514 × 768
30	LayerNorm_7_1	LayerNorm	1 × 514 × 768
31	FFN_7	Feed Forward	1 × 514 × 768
32	Attention_8	Multi-Head Attention	1 × 514 × 768
33	Add_8	Add	1 × 514 × 768
34	LayerNorm_8_1	LayerNorm	1 × 514 × 768
35	FFN_8	Feed Forward	1 × 514 × 768
36	Attention_9	Multi-Head Attention	1 × 514 × 768
37	Add_9	Add	1 × 514 × 768
38	LayerNorm_9_1	LayerNorm	1 × 514 × 768
39	FFN_9	Feed Forward	1 × 514 × 768
40	Attention_10	Multi-Head Attention	1 × 514 × 768
41	Add_10	Add	1 × 514 × 768
42	LayerNorm_10_1	LayerNorm	1 × 514 × 768
43	FFN_10	Feed Forward	1 × 514 × 768
44	Attention_11	Multi-Head Attention	1 × 514 × 768
45	Add_11	Add	1 × 514 × 768
46	LayerNorm_11_1	LayerNorm	1 × 514 × 768
47	FFN_11	Feed Forward	1 × 514 × 768
48	Attention_12	Multi-Head Attention	1 × 514 × 768
49	Add_12	Add	1 × 514 × 768
50	LayerNorm_12_1	LayerNorm	1 × 514 × 768
51	FFN_12	Feed Forward	1 × 514 × 768
52	Output	Output	1 × 514 × 768

What the verifier says

infoAt 12 stacked attention layers, residual-branch outputs add up; unscaled init lets activation variance grow with depth. GPT-2/LLaMA-family models scale the residual projections by depth (N(0, 0.02 / √(2L))). Fix: Scale residual output projections by depth: nn.init.normal_(w, std=0.02 / math.sqrt(2 * n_layers))
deep-attention-default-init

Do this to your own model

Same numbers, on a model in your repo, in one command. No account.

pip install neurarch-trace
neurarch-trace FacebookAI/roberta-base --plan --share

Paste a model instead Markdown twin Index JSON POST /api/v1/plan