36 1 week ago

A tiny 150M completion model trained from scratch for short story generation and small-model pipeline validation. It is best for generating small story prompts. (Think young child telling small stories. And cute at times.)

150m
2524261bad58 · 269MB
    Metadata
  • general.architecture
    llama
  • general.file_type
    F16
  • llama.attention.head_count
    12
  • llama.attention.head_count_kv
    12
  • llama.attention.key_length
    64
  • llama.attention.layer_norm_rms_epsilon
    1e-06
  • llama.attention.value_length
    64
  • llama.block_count
    12
  • llama.context_length
    1024
  • llama.embedding_length
    768
  • llama.feed_forward_length
    2048
  • llama.rope.dimension_count
    64
  • llama.rope.freq_base
    10000
  • llama.vocab_size
    32000
  • tokenizer.ggml.add_bos_token
    true
  • tokenizer.ggml.add_eos_token
    false
  • tokenizer.ggml.add_space_prefix
    true
  • tokenizer.ggml.bos_token_id
    1
  • tokenizer.ggml.eos_token_id
    2
  • tokenizer.ggml.model
    llama
  • tokenizer.ggml.padding_token_id
    0
  • tokenizer.ggml.pre
    default
  • tokenizer.ggml.scores
    [-1000, -1000, -1000, -1000, -0, ...]
  • tokenizer.ggml.token_type
    [3, 3, 3, 3, 1, ...]
  • tokenizer.ggml.tokens
    [<pad>, <s>, </s>, <unk>, he, ...]
  • tokenizer.ggml.unknown_token_id
    3
  • Tensor
  • token_embd.weight
    F16
    [768, 32000]
  • blk.0
  • blk.0.attn_k.weight
    F16
    [768, 768]
  • blk.0.attn_norm.weight
    F32
    [768]
  • blk.0.attn_output.weight
    F16
    [768, 768]
  • blk.0.attn_q.weight
    F16
    [768, 768]
  • blk.0.attn_v.weight
    F16
    [768, 768]
  • blk.0.ffn_down.weight
    F16
    [2048, 768]
  • blk.0.ffn_gate.weight
    F16
    [768, 2048]
  • blk.0.ffn_norm.weight
    F32
    [768]
  • blk.0.ffn_up.weight
    F16
    [768, 2048]
  • blk.1
  • blk.1.attn_k.weight
    F16
    [768, 768]
  • blk.1.attn_norm.weight
    F32
    [768]
  • blk.1.attn_output.weight
    F16
    [768, 768]
  • blk.1.attn_q.weight
    F16
    [768, 768]
  • blk.1.attn_v.weight
    F16
    [768, 768]
  • blk.1.ffn_down.weight
    F16
    [2048, 768]
  • blk.1.ffn_gate.weight
    F16
    [768, 2048]
  • blk.1.ffn_norm.weight
    F32
    [768]
  • blk.1.ffn_up.weight
    F16
    [768, 2048]
  • blk.2
  • blk.2.attn_k.weight
    F16
    [768, 768]
  • blk.2.attn_norm.weight
    F32
    [768]
  • blk.2.attn_output.weight
    F16
    [768, 768]
  • blk.2.attn_q.weight
    F16
    [768, 768]
  • blk.2.attn_v.weight
    F16
    [768, 768]
  • blk.2.ffn_down.weight
    F16
    [2048, 768]
  • blk.2.ffn_gate.weight
    F16
    [768, 2048]
  • blk.2.ffn_norm.weight
    F32
    [768]
  • blk.2.ffn_up.weight
    F16
    [768, 2048]
  • blk.3
  • blk.3.attn_k.weight
    F16
    [768, 768]
  • blk.3.attn_norm.weight
    F32
    [768]
  • blk.3.attn_output.weight
    F16
    [768, 768]
  • blk.3.attn_q.weight
    F16
    [768, 768]
  • blk.3.attn_v.weight
    F16
    [768, 768]
  • blk.3.ffn_down.weight
    F16
    [2048, 768]
  • blk.3.ffn_gate.weight
    F16
    [768, 2048]
  • blk.3.ffn_norm.weight
    F32
    [768]
  • blk.3.ffn_up.weight
    F16
    [768, 2048]
  • blk.4
  • blk.4.attn_k.weight
    F16
    [768, 768]
  • blk.4.attn_norm.weight
    F32
    [768]
  • blk.4.attn_output.weight
    F16
    [768, 768]
  • blk.4.attn_q.weight
    F16
    [768, 768]
  • blk.4.attn_v.weight
    F16
    [768, 768]
  • blk.4.ffn_down.weight
    F16
    [2048, 768]
  • blk.4.ffn_gate.weight
    F16
    [768, 2048]
  • blk.4.ffn_norm.weight
    F32
    [768]
  • blk.4.ffn_up.weight
    F16
    [768, 2048]
  • blk.5
  • blk.5.attn_k.weight
    F16
    [768, 768]
  • blk.5.attn_norm.weight
    F32
    [768]
  • blk.5.attn_output.weight
    F16
    [768, 768]
  • blk.5.attn_q.weight
    F16
    [768, 768]
  • blk.5.attn_v.weight
    F16
    [768, 768]
  • blk.5.ffn_down.weight
    F16
    [2048, 768]
  • blk.5.ffn_gate.weight
    F16
    [768, 2048]
  • blk.5.ffn_norm.weight
    F32
    [768]
  • blk.5.ffn_up.weight
    F16
    [768, 2048]
  • blk.6
  • blk.6.attn_k.weight
    F16
    [768, 768]
  • blk.6.attn_norm.weight
    F32
    [768]
  • blk.6.attn_output.weight
    F16
    [768, 768]
  • blk.6.attn_q.weight
    F16
    [768, 768]
  • blk.6.attn_v.weight
    F16
    [768, 768]
  • blk.6.ffn_down.weight
    F16
    [2048, 768]
  • blk.6.ffn_gate.weight
    F16
    [768, 2048]
  • blk.6.ffn_norm.weight
    F32
    [768]
  • blk.6.ffn_up.weight
    F16
    [768, 2048]
  • blk.7
  • blk.7.attn_k.weight
    F16
    [768, 768]
  • blk.7.attn_norm.weight
    F32
    [768]
  • blk.7.attn_output.weight
    F16
    [768, 768]
  • blk.7.attn_q.weight
    F16
    [768, 768]
  • blk.7.attn_v.weight
    F16
    [768, 768]
  • blk.7.ffn_down.weight
    F16
    [2048, 768]
  • blk.7.ffn_gate.weight
    F16
    [768, 2048]
  • blk.7.ffn_norm.weight
    F32
    [768]
  • blk.7.ffn_up.weight
    F16
    [768, 2048]
  • blk.8
  • blk.8.attn_k.weight
    F16
    [768, 768]
  • blk.8.attn_norm.weight
    F32
    [768]
  • blk.8.attn_output.weight
    F16
    [768, 768]
  • blk.8.attn_q.weight
    F16
    [768, 768]
  • blk.8.attn_v.weight
    F16
    [768, 768]
  • blk.8.ffn_down.weight
    F16
    [2048, 768]
  • blk.8.ffn_gate.weight
    F16
    [768, 2048]
  • blk.8.ffn_norm.weight
    F32
    [768]
  • blk.8.ffn_up.weight
    F16
    [768, 2048]
  • blk.9
  • blk.9.attn_k.weight
    F16
    [768, 768]
  • blk.9.attn_norm.weight
    F32
    [768]
  • blk.9.attn_output.weight
    F16
    [768, 768]
  • blk.9.attn_q.weight
    F16
    [768, 768]
  • blk.9.attn_v.weight
    F16
    [768, 768]
  • blk.9.ffn_down.weight
    F16
    [2048, 768]
  • blk.9.ffn_gate.weight
    F16
    [768, 2048]
  • blk.9.ffn_norm.weight
    F32
    [768]
  • blk.9.ffn_up.weight
    F16
    [768, 2048]
  • blk.10
  • blk.10.attn_k.weight
    F16
    [768, 768]
  • blk.10.attn_norm.weight
    F32
    [768]
  • blk.10.attn_output.weight
    F16
    [768, 768]
  • blk.10.attn_q.weight
    F16
    [768, 768]
  • blk.10.attn_v.weight
    F16
    [768, 768]
  • blk.10.ffn_down.weight
    F16
    [2048, 768]
  • blk.10.ffn_gate.weight
    F16
    [768, 2048]
  • blk.10.ffn_norm.weight
    F32
    [768]
  • blk.10.ffn_up.weight
    F16
    [768, 2048]
  • blk.11
  • blk.11.attn_k.weight
    F16
    [768, 768]
  • blk.11.attn_norm.weight
    F32
    [768]
  • blk.11.attn_output.weight
    F16
    [768, 768]
  • blk.11.attn_q.weight
    F16
    [768, 768]
  • blk.11.attn_v.weight
    F16
    [768, 768]
  • blk.11.ffn_down.weight
    F16
    [2048, 768]
  • blk.11.ffn_gate.weight
    F16
    [768, 2048]
  • blk.11.ffn_norm.weight
    F32
    [768]
  • blk.11.ffn_up.weight
    F16
    [768, 2048]
  • output.weight
    F16
    [768, 32000]
  • output_norm.weight
    F32
    [768]