whereisanzi/ maracatu-20m:latest

9 1 month ago

17M-param Brazilian Portuguese LLM, trained from scratch on Wikipedia PT. Base model, Apache 2.0.

... /
model
f9c04e799b55 · 14MB
    Metadata
  • general.architecture
    llama
  • general.file_type
    Q4_K_M
  • llama.attention.head_count
    6
  • llama.attention.head_count_kv
    6
  • llama.attention.key_length
    64
  • llama.attention.layer_norm_rms_epsilon
    1e-05
  • llama.attention.value_length
    64
  • llama.block_count
    6
  • llama.context_length
    512
  • llama.embedding_length
    384
  • llama.feed_forward_length
    1024
  • llama.rope.dimension_count
    64
  • llama.rope.freq_base
    10000
  • llama.vocab_size
    16000
  • tokenizer.ggml.add_bos_token
    true
  • tokenizer.ggml.add_eos_token
    false
  • tokenizer.ggml.bos_token_id
    2
  • tokenizer.ggml.eos_token_id
    3
  • tokenizer.ggml.model
    llama
  • tokenizer.ggml.padding_token_id
    0
  • tokenizer.ggml.pre
    default
  • tokenizer.ggml.scores
    [0, 0, 0, 0, 0, ...]
  • tokenizer.ggml.token_type
    [3, 2, 3, 3, 6, ...]
  • tokenizer.ggml.tokens
    [<pad>, <unk>, <s>, </s>, <0x00>, ...]
    • Tensor
  • token_embd.weight
    Q8_0
    [384, 16000]
    • blk.0
  • blk.0.attn_k.weight
    Q5_0
    [384, 384]
  • blk.0.attn_norm.weight
    F32
    [384]
  • blk.0.attn_output.weight
    Q5_0
    [384, 384]
  • blk.0.attn_q.weight
    Q5_0
    [384, 384]
  • blk.0.attn_v.weight
    Q5_0
    [384, 384]
  • blk.0.ffn_down.weight
    Q4_K
    [1024, 384]
  • blk.0.ffn_gate.weight
    Q5_0
    [384, 1024]
  • blk.0.ffn_norm.weight
    F32
    [384]
  • blk.0.ffn_up.weight
    Q5_0
    [384, 1024]
    • blk.1
  • blk.1.attn_k.weight
    Q5_0
    [384, 384]
  • blk.1.attn_norm.weight
    F32
    [384]
  • blk.1.attn_output.weight
    Q5_0
    [384, 384]
  • blk.1.attn_q.weight
    Q5_0
    [384, 384]
  • blk.1.attn_v.weight
    Q5_0
    [384, 384]
  • blk.1.ffn_down.weight
    Q4_K
    [1024, 384]
  • blk.1.ffn_gate.weight
    Q5_0
    [384, 1024]
  • blk.1.ffn_norm.weight
    F32
    [384]
  • blk.1.ffn_up.weight
    Q5_0
    [384, 1024]
    • blk.2
  • blk.2.attn_k.weight
    Q5_0
    [384, 384]
  • blk.2.attn_norm.weight
    F32
    [384]
  • blk.2.attn_output.weight
    Q5_0
    [384, 384]
  • blk.2.attn_q.weight
    Q5_0
    [384, 384]
  • blk.2.attn_v.weight
    Q8_0
    [384, 384]
  • blk.2.ffn_down.weight
    Q6_K
    [1024, 384]
  • blk.2.ffn_gate.weight
    Q5_0
    [384, 1024]
  • blk.2.ffn_norm.weight
    F32
    [384]
  • blk.2.ffn_up.weight
    Q5_0
    [384, 1024]
    • blk.3
  • blk.3.attn_k.weight
    Q5_0
    [384, 384]
  • blk.3.attn_norm.weight
    F32
    [384]
  • blk.3.attn_output.weight
    Q5_0
    [384, 384]
  • blk.3.attn_q.weight
    Q5_0
    [384, 384]
  • blk.3.attn_v.weight
    Q5_0
    [384, 384]
  • blk.3.ffn_down.weight
    Q4_K
    [1024, 384]
  • blk.3.ffn_gate.weight
    Q5_0
    [384, 1024]
  • blk.3.ffn_norm.weight
    F32
    [384]
  • blk.3.ffn_up.weight
    Q5_0
    [384, 1024]
    • blk.4
  • blk.4.attn_k.weight
    Q5_0
    [384, 384]
  • blk.4.attn_norm.weight
    F32
    [384]
  • blk.4.attn_output.weight
    Q5_0
    [384, 384]
  • blk.4.attn_q.weight
    Q5_0
    [384, 384]
  • blk.4.attn_v.weight
    Q5_0
    [384, 384]
  • blk.4.ffn_down.weight
    Q4_K
    [1024, 384]
  • blk.4.ffn_gate.weight
    Q5_0
    [384, 1024]
  • blk.4.ffn_norm.weight
    F32
    [384]
  • blk.4.ffn_up.weight
    Q5_0
    [384, 1024]
    • blk.5
  • blk.5.attn_k.weight
    Q5_0
    [384, 384]
  • blk.5.attn_norm.weight
    F32
    [384]
  • blk.5.attn_output.weight
    Q5_0
    [384, 384]
  • blk.5.attn_q.weight
    Q5_0
    [384, 384]
  • blk.5.attn_v.weight
    Q8_0
    [384, 384]
  • blk.5.ffn_down.weight
    Q6_K
    [1024, 384]
  • blk.5.ffn_gate.weight
    Q5_0
    [384, 1024]
  • blk.5.ffn_norm.weight
    F32
    [384]
  • blk.5.ffn_up.weight
    Q5_0
    [384, 1024]
  • output_norm.weight
    F32
    [384]