JXW67/TGAI_NB:0.86b/model

JXW67/ TGAI_NB:0.86b

21 Downloads Updated 1 week ago

TGAI NB — a lightweight Chinese MoE chat model family, built end-to-end by a high school student. V2 (3B): 8-expert sparse MoE, better quality for desktop. V1 (0.86B): 4-expert sparse MoE, ultra-light, smooth on low-end devices.

0.86b 3b

TGAI_NB:0.86b ... /

model

8f42636ebf6c · 1.7GB

Metadata

general.architecture

llama

llama
general.file_type

F16

F16
llama.attention.head_count

14

14
llama.attention.head_count_kv

14

14
llama.attention.layer_norm_rms_epsilon

1e-06

1e-06
llama.block_count

20

20
llama.context_length

512

512
llama.embedding_length

896

896
llama.expert_count

4

4
llama.expert_used_count

2

2
llama.feed_forward_length

3584

3584
llama.rope.freq_base

10000

10000
llama.vocab_size

8668

8668
tokenizer.ggml.add_bos_token

true

true
tokenizer.ggml.add_eos_token

false

false
tokenizer.ggml.add_space_prefix

false

false
tokenizer.ggml.bos_token_id

2

2
tokenizer.ggml.byte_fallback

false

false
tokenizer.ggml.eos_token_id

3

3
tokenizer.ggml.merges

[]

[]
tokenizer.ggml.model

llama

llama
tokenizer.ggml.scores

[0, 0, 0, 0, 0, ...]

[0, 0, 0, 0, 0, ...]
tokenizer.ggml.tokens

[<PAD>, <UNK>, <BOS>, <EOS>, <IMG>, ...]

[<PAD>, <UNK>, <BOS>, <EOS>, <IMG>, ...]
tokenizer.ggml.unk_token_id

1

1

Tensor

Name

Type

Shape
token_embd.weight

F16

F16

[896, 8668]

blk.0

blk.0.attn_k.weight

F16

F16

[896, 896]
blk.0.attn_norm.weight

F32

F32

[896]
blk.0.attn_output.weight

F16

F16

[896, 896]
blk.0.attn_q.weight

F16

F16

[896, 896]
blk.0.attn_v.weight

F16

F16

[896, 896]
blk.0.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.0.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.0.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.0.ffn_norm.weight

F32

F32

[896]
blk.0.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.1

blk.1.attn_k.weight

F16

F16

[896, 896]
blk.1.attn_norm.weight

F32

F32

[896]
blk.1.attn_output.weight

F16

F16

[896, 896]
blk.1.attn_q.weight

F16

F16

[896, 896]
blk.1.attn_v.weight

F16

F16

[896, 896]
blk.1.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.1.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.1.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.1.ffn_norm.weight

F32

F32

[896]
blk.1.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.2

blk.2.attn_k.weight

F16

F16

[896, 896]
blk.2.attn_norm.weight

F32

F32

[896]
blk.2.attn_output.weight

F16

F16

[896, 896]
blk.2.attn_q.weight

F16

F16

[896, 896]
blk.2.attn_v.weight

F16

F16

[896, 896]
blk.2.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.2.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.2.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.2.ffn_norm.weight

F32

F32

[896]
blk.2.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.3

blk.3.attn_k.weight

F16

F16

[896, 896]
blk.3.attn_norm.weight

F32

F32

[896]
blk.3.attn_output.weight

F16

F16

[896, 896]
blk.3.attn_q.weight

F16

F16

[896, 896]
blk.3.attn_v.weight

F16

F16

[896, 896]
blk.3.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.3.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.3.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.3.ffn_norm.weight

F32

F32

[896]
blk.3.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.4

blk.4.attn_k.weight

F16

F16

[896, 896]
blk.4.attn_norm.weight

F32

F32

[896]
blk.4.attn_output.weight

F16

F16

[896, 896]
blk.4.attn_q.weight

F16

F16

[896, 896]
blk.4.attn_v.weight

F16

F16

[896, 896]
blk.4.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.4.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.4.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.4.ffn_norm.weight

F32

F32

[896]
blk.4.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.5

blk.5.attn_k.weight

F16

F16

[896, 896]
blk.5.attn_norm.weight

F32

F32

[896]
blk.5.attn_output.weight

F16

F16

[896, 896]
blk.5.attn_q.weight

F16

F16

[896, 896]
blk.5.attn_v.weight

F16

F16

[896, 896]
blk.5.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.5.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.5.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.5.ffn_norm.weight

F32

F32

[896]
blk.5.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.6

blk.6.attn_k.weight

F16

F16

[896, 896]
blk.6.attn_norm.weight

F32

F32

[896]
blk.6.attn_output.weight

F16

F16

[896, 896]
blk.6.attn_q.weight

F16

F16

[896, 896]
blk.6.attn_v.weight

F16

F16

[896, 896]
blk.6.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.6.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.6.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.6.ffn_norm.weight

F32

F32

[896]
blk.6.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.7

blk.7.attn_k.weight

F16

F16

[896, 896]
blk.7.attn_norm.weight

F32

F32

[896]
blk.7.attn_output.weight

F16

F16

[896, 896]
blk.7.attn_q.weight

F16

F16

[896, 896]
blk.7.attn_v.weight

F16

F16

[896, 896]
blk.7.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.7.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.7.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.7.ffn_norm.weight

F32

F32

[896]
blk.7.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.8

blk.8.attn_k.weight

F16

F16

[896, 896]
blk.8.attn_norm.weight

F32

F32

[896]
blk.8.attn_output.weight

F16

F16

[896, 896]
blk.8.attn_q.weight

F16

F16

[896, 896]
blk.8.attn_v.weight

F16

F16

[896, 896]
blk.8.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.8.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.8.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.8.ffn_norm.weight

F32

F32

[896]
blk.8.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.9

blk.9.attn_k.weight

F16

F16

[896, 896]
blk.9.attn_norm.weight

F32

F32

[896]
blk.9.attn_output.weight

F16

F16

[896, 896]
blk.9.attn_q.weight

F16

F16

[896, 896]
blk.9.attn_v.weight

F16

F16

[896, 896]
blk.9.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.9.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.9.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.9.ffn_norm.weight

F32

F32

[896]
blk.9.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.10

blk.10.attn_k.weight

F16

F16

[896, 896]
blk.10.attn_norm.weight

F32

F32

[896]
blk.10.attn_output.weight

F16

F16

[896, 896]
blk.10.attn_q.weight

F16

F16

[896, 896]
blk.10.attn_v.weight

F16

F16

[896, 896]
blk.10.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.10.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.10.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.10.ffn_norm.weight

F32

F32

[896]
blk.10.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.11

blk.11.attn_k.weight

F16

F16

[896, 896]
blk.11.attn_norm.weight

F32

F32

[896]
blk.11.attn_output.weight

F16

F16

[896, 896]
blk.11.attn_q.weight

F16

F16

[896, 896]
blk.11.attn_v.weight

F16

F16

[896, 896]
blk.11.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.11.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.11.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.11.ffn_norm.weight

F32

F32

[896]
blk.11.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.12

blk.12.attn_k.weight

F16

F16

[896, 896]
blk.12.attn_norm.weight

F32

F32

[896]
blk.12.attn_output.weight

F16

F16

[896, 896]
blk.12.attn_q.weight

F16

F16

[896, 896]
blk.12.attn_v.weight

F16

F16

[896, 896]
blk.12.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.12.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.12.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.12.ffn_norm.weight

F32

F32

[896]
blk.12.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.13

blk.13.attn_k.weight

F16

F16

[896, 896]
blk.13.attn_norm.weight

F32

F32

[896]
blk.13.attn_output.weight

F16

F16

[896, 896]
blk.13.attn_q.weight

F16

F16

[896, 896]
blk.13.attn_v.weight

F16

F16

[896, 896]
blk.13.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.13.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.13.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.13.ffn_norm.weight

F32

F32

[896]
blk.13.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.14

blk.14.attn_k.weight

F16

F16

[896, 896]
blk.14.attn_norm.weight

F32

F32

[896]
blk.14.attn_output.weight

F16

F16

[896, 896]
blk.14.attn_q.weight

F16

F16

[896, 896]
blk.14.attn_v.weight

F16

F16

[896, 896]
blk.14.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.14.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.14.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.14.ffn_norm.weight

F32

F32

[896]
blk.14.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.15

blk.15.attn_k.weight

F16

F16

[896, 896]
blk.15.attn_norm.weight

F32

F32

[896]
blk.15.attn_output.weight

F16

F16

[896, 896]
blk.15.attn_q.weight

F16

F16

[896, 896]
blk.15.attn_v.weight

F16

F16

[896, 896]
blk.15.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.15.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.15.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.15.ffn_norm.weight

F32

F32

[896]
blk.15.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.16

blk.16.attn_k.weight

F16

F16

[896, 896]
blk.16.attn_norm.weight

F32

F32

[896]
blk.16.attn_output.weight

F16

F16

[896, 896]
blk.16.attn_q.weight

F16

F16

[896, 896]
blk.16.attn_v.weight

F16

F16

[896, 896]
blk.16.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.16.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.16.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.16.ffn_norm.weight

F32

F32

[896]
blk.16.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.17

blk.17.attn_k.weight

F16

F16

[896, 896]
blk.17.attn_norm.weight

F32

F32

[896]
blk.17.attn_output.weight

F16

F16

[896, 896]
blk.17.attn_q.weight

F16

F16

[896, 896]
blk.17.attn_v.weight

F16

F16

[896, 896]
blk.17.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.17.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.17.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.17.ffn_norm.weight

F32

F32

[896]
blk.17.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.18

blk.18.attn_k.weight

F16

F16

[896, 896]
blk.18.attn_norm.weight

F32

F32

[896]
blk.18.attn_output.weight

F16

F16

[896, 896]
blk.18.attn_q.weight

F16

F16

[896, 896]
blk.18.attn_v.weight

F16

F16

[896, 896]
blk.18.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.18.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.18.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.18.ffn_norm.weight

F32

F32

[896]
blk.18.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

blk.19

blk.19.attn_k.weight

F16

F16

[896, 896]
blk.19.attn_norm.weight

F32

F32

[896]
blk.19.attn_output.weight

F16

F16

[896, 896]
blk.19.attn_q.weight

F16

F16

[896, 896]
blk.19.attn_v.weight

F16

F16

[896, 896]
blk.19.ffn_down_exps.weight

F16

F16

[3584, 896, 4]
blk.19.ffn_gate_exps.weight

F16

F16

[896, 3584, 4]
blk.19.ffn_gate_inp.weight

F16

F16

[896, 4]
blk.19.ffn_norm.weight

F32

F32

[896]
blk.19.ffn_up_exps.weight

F16

F16

[896, 3584, 4]

output_norm.weight

F32

F32

[896]