← MIC coverage of Transformers model types

google/bigbird-roberta-base

huggingface-mic/bigbird-roberta-base-mic@5c8d7c34dc91 ← google/bigbird-roberta-base@5a145f7852cb · BigBirdForPreTraining · random weights

CI: no failure on main in 30 days · MIC failures dashboard

Original128.1Mparameters
Fixture3.6Mparameters
Layer stacks1compared
Findings0to review

✓ Same architecture, smaller

Every module class, layer kind and checked property of the original is in the fixture.

Layer stacks

One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.

bert.encoder.layer

Original · 12
Fixture · 2
KindOriginalFixture
Layer122✓ kept

Where the parameters are

Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.

Original
Fixture
ComponentOriginalShareFixtureShare
bert.encoder.layer85.1M66.4%66.9K1.9%
bert.embeddings41.8M32.6%3.5M96.5%
cls.predictions642.5K0.5%54.6K1.5%
bert.pooler590.6K0.5%4.2K0.1%
cls.seq_relationship1.5K0.0%1300.0%

Preserved properties

ConfigPropertyOriginalFixture
modelmodel_typebig_birdbig_bird✓ kept
modelvocab_size5035850358✓ kept
modeltie_word_embeddingsTrueTrue✓ kept
modelhidden_actgelu_newgelu_new✓ kept
modelmax_position_embeddings40964096✓ kept

Module classes

Instances of each class in the built model (prefix BigBird omitted).

ClassOriginalFixture
Linear7616
LayerNorm266
Dropout255
NewGELUActivation133
Attention122
BlockSparseAttention122
Intermediate122
Layer122
Output122
SelfOutput122
Embedding33
Embeddings11
Encoder11
ForPreTraining11
LMPredictionHead11
Model11
PreTrainingHeads11
PredictionHeadTransform11
ModuleList11
Tanh11

Config changes

4 fields changed, 33 unchanged.

FieldOriginalFixture
hidden_size76864
intermediate_size3072128
num_attention_heads124
num_hidden_layers122