rafi-dev/rb-nano/model

rafi-dev/ rb-nano:latest

9 Downloads Updated 2 days ago

rb-nano is a 48M-parameter chat model trained from scratch on a single consumer GPU — tiny, fast, and CPU-friendly.

rb-nano:latest ... /

model

41f048980a0b · 131MB

Metadata

general.architecture

gpt2

gpt2
general.file_type

F16

F16
gpt2.attention.head_count

8

8
gpt2.attention.layer_norm_epsilon

1e-05

1e-05
gpt2.block_count

10

10
gpt2.context_length

1024

1024
gpt2.embedding_length

512

512
gpt2.feed_forward_length

2048

2048
tokenizer.ggml.add_bos_token

true

true
tokenizer.ggml.bos_token_id

1

1
tokenizer.ggml.eos_token_id

2

2
tokenizer.ggml.merges

[Ġ t, Ġ a, h e, i n, r e, ...]

[Ġ t, Ġ a, h e, i n, r e, ...]
tokenizer.ggml.model

gpt2

gpt2
tokenizer.ggml.padding_token_id

0

0
tokenizer.ggml.pre

default

default
tokenizer.ggml.token_type

[3, 3, 3, 3, 1, ...]

[3, 3, 3, 3, 1, ...]
tokenizer.ggml.tokens

[<pad>, <sos>, <eos>, <unk>, !, ...]

[<pad>, <sos>, <eos>, <unk>, !, ...]
tokenizer.ggml.unknown_token_id

3

3

Tensor

Name

Type

Shape
token_embd.weight

F16

F16

[512, 32000]

blk.0

blk.0.attn_norm.bias

F32

F32

[512]
blk.0.attn_norm.weight

F32

F32

[512]
blk.0.attn_output.bias

F32

F32

[512]
blk.0.attn_output.weight

F16

F16

[512, 512]
blk.0.attn_qkv.bias

F32

F32

[1536]
blk.0.attn_qkv.weight

F16

F16

[512, 1536]
blk.0.ffn_down.bias

F32

F32

[512]
blk.0.ffn_down.weight

F16

F16

[2048, 512]
blk.0.ffn_norm.bias

F32

F32

[512]
blk.0.ffn_norm.weight

F32

F32

[512]
blk.0.ffn_up.bias

F32

F32

[2048]
blk.0.ffn_up.weight

F16

F16

[512, 2048]

blk.1

blk.1.attn_norm.bias

F32

F32

[512]
blk.1.attn_norm.weight

F32

F32

[512]
blk.1.attn_output.bias

F32

F32

[512]
blk.1.attn_output.weight

F16

F16

[512, 512]
blk.1.attn_qkv.bias

F32

F32

[1536]
blk.1.attn_qkv.weight

F16

F16

[512, 1536]
blk.1.ffn_down.bias

F32

F32

[512]
blk.1.ffn_down.weight

F16

F16

[2048, 512]
blk.1.ffn_norm.bias

F32

F32

[512]
blk.1.ffn_norm.weight

F32

F32

[512]
blk.1.ffn_up.bias

F32

F32

[2048]
blk.1.ffn_up.weight

F16

F16

[512, 2048]

blk.2

blk.2.attn_norm.bias

F32

F32

[512]
blk.2.attn_norm.weight

F32

F32

[512]
blk.2.attn_output.bias

F32

F32

[512]
blk.2.attn_output.weight

F16

F16

[512, 512]
blk.2.attn_qkv.bias

F32

F32

[1536]
blk.2.attn_qkv.weight

F16

F16

[512, 1536]
blk.2.ffn_down.bias

F32

F32

[512]
blk.2.ffn_down.weight

F16

F16

[2048, 512]
blk.2.ffn_norm.bias

F32

F32

[512]
blk.2.ffn_norm.weight

F32

F32

[512]
blk.2.ffn_up.bias

F32

F32

[2048]
blk.2.ffn_up.weight

F16

F16

[512, 2048]

blk.3

blk.3.attn_norm.bias

F32

F32

[512]
blk.3.attn_norm.weight

F32

F32

[512]
blk.3.attn_output.bias

F32

F32

[512]
blk.3.attn_output.weight

F16

F16

[512, 512]
blk.3.attn_qkv.bias

F32

F32

[1536]
blk.3.attn_qkv.weight

F16

F16

[512, 1536]
blk.3.ffn_down.bias

F32

F32

[512]
blk.3.ffn_down.weight

F16

F16

[2048, 512]
blk.3.ffn_norm.bias

F32

F32

[512]
blk.3.ffn_norm.weight

F32

F32

[512]
blk.3.ffn_up.bias

F32

F32

[2048]
blk.3.ffn_up.weight

F16

F16

[512, 2048]

blk.4

blk.4.attn_norm.bias

F32

F32

[512]
blk.4.attn_norm.weight

F32

F32

[512]
blk.4.attn_output.bias

F32

F32

[512]
blk.4.attn_output.weight

F16

F16

[512, 512]
blk.4.attn_qkv.bias

F32

F32

[1536]
blk.4.attn_qkv.weight

F16

F16

[512, 1536]
blk.4.ffn_down.bias

F32

F32

[512]
blk.4.ffn_down.weight

F16

F16

[2048, 512]
blk.4.ffn_norm.bias

F32

F32

[512]
blk.4.ffn_norm.weight

F32

F32

[512]
blk.4.ffn_up.bias

F32

F32

[2048]
blk.4.ffn_up.weight

F16

F16

[512, 2048]

blk.5

blk.5.attn_norm.bias

F32

F32

[512]
blk.5.attn_norm.weight

F32

F32

[512]
blk.5.attn_output.bias

F32

F32

[512]
blk.5.attn_output.weight

F16

F16

[512, 512]
blk.5.attn_qkv.bias

F32

F32

[1536]
blk.5.attn_qkv.weight

F16

F16

[512, 1536]
blk.5.ffn_down.bias

F32

F32

[512]
blk.5.ffn_down.weight

F16

F16

[2048, 512]
blk.5.ffn_norm.bias

F32

F32

[512]
blk.5.ffn_norm.weight

F32

F32

[512]
blk.5.ffn_up.bias

F32

F32

[2048]
blk.5.ffn_up.weight

F16

F16

[512, 2048]

blk.6

blk.6.attn_norm.bias

F32

F32

[512]
blk.6.attn_norm.weight

F32

F32

[512]
blk.6.attn_output.bias

F32

F32

[512]
blk.6.attn_output.weight

F16

F16

[512, 512]
blk.6.attn_qkv.bias

F32

F32

[1536]
blk.6.attn_qkv.weight

F16

F16

[512, 1536]
blk.6.ffn_down.bias

F32

F32

[512]
blk.6.ffn_down.weight

F16

F16

[2048, 512]
blk.6.ffn_norm.bias

F32

F32

[512]
blk.6.ffn_norm.weight

F32

F32

[512]
blk.6.ffn_up.bias

F32

F32

[2048]
blk.6.ffn_up.weight

F16

F16

[512, 2048]

blk.7

blk.7.attn_norm.bias

F32

F32

[512]
blk.7.attn_norm.weight

F32

F32

[512]
blk.7.attn_output.bias

F32

F32

[512]
blk.7.attn_output.weight

F16

F16

[512, 512]
blk.7.attn_qkv.bias

F32

F32

[1536]
blk.7.attn_qkv.weight

F16

F16

[512, 1536]
blk.7.ffn_down.bias

F32

F32

[512]
blk.7.ffn_down.weight

F16

F16

[2048, 512]
blk.7.ffn_norm.bias

F32

F32

[512]
blk.7.ffn_norm.weight

F32

F32

[512]
blk.7.ffn_up.bias

F32

F32

[2048]
blk.7.ffn_up.weight

F16

F16

[512, 2048]

blk.8

blk.8.attn_norm.bias

F32

F32

[512]
blk.8.attn_norm.weight

F32

F32

[512]
blk.8.attn_output.bias

F32

F32

[512]
blk.8.attn_output.weight

F16

F16

[512, 512]
blk.8.attn_qkv.bias

F32

F32

[1536]
blk.8.attn_qkv.weight

F16

F16

[512, 1536]
blk.8.ffn_down.bias

F32

F32

[512]
blk.8.ffn_down.weight

F16

F16

[2048, 512]
blk.8.ffn_norm.bias

F32

F32

[512]
blk.8.ffn_norm.weight

F32

F32

[512]
blk.8.ffn_up.bias

F32

F32

[2048]
blk.8.ffn_up.weight

F16

F16

[512, 2048]

blk.9

blk.9.attn_norm.bias

F32

F32

[512]
blk.9.attn_norm.weight

F32

F32

[512]
blk.9.attn_output.bias

F32

F32

[512]
blk.9.attn_output.weight

F16

F16

[512, 512]
blk.9.attn_qkv.bias

F32

F32

[1536]
blk.9.attn_qkv.weight

F16

F16

[512, 1536]
blk.9.ffn_down.bias

F32

F32

[512]
blk.9.ffn_down.weight

F16

F16

[2048, 512]
blk.9.ffn_norm.bias

F32

F32

[512]
blk.9.ffn_norm.weight

F32

F32

[512]
blk.9.ffn_up.bias

F32

F32

[2048]
blk.9.ffn_up.weight

F16

F16

[512, 2048]

output.weight

F16

F16

[512, 32000]
output_norm.bias

F32

F32

[512]
position_embd.weight

F16

F16

[512, 1024]
output_norm.weight

F32

F32

[512]