sppandita85/ story-llm:latest

16 1 week ago

A GPT-style language model built entirely from scratch (no HuggingFace/nanoGPT for the model itself) and trained on 50 children's moral stories. It's an educational reproduction of the full LLM lifecycle at miniature scale.

... /
model
2db06ac8b8ab · 2.0MB
    Metadata
  • general.architecture
    gpt2
  • general.file_type
    F16
  • gpt2.attention.head_count
    4
  • gpt2.attention.layer_norm_epsilon
    1e-05
  • gpt2.block_count
    4
  • gpt2.context_length
    128
  • gpt2.embedding_length
    128
  • gpt2.feed_forward_length
    512
  • tokenizer.ggml.add_bos_token
    false
  • tokenizer.ggml.bos_token_id
    1020
  • tokenizer.ggml.eos_token_id
    1020
  • tokenizer.ggml.merges
    [h e, Ġ t, Ġ a, i n, e d, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    1023
  • tokenizer.ggml.pre
    gpt-2
  • tokenizer.ggml.token_type
    [1, 1, 1, 1, 1, ...]
  • tokenizer.ggml.tokens
    [Ā, ā, Ă, ă, Ą, ...]
  • tokenizer.ggml.unknown_token_id
    1020
    • Tensor
  • token_embd.weight
    F16
    [128, 1024]
    • blk.0
  • blk.0.attn_norm.bias
    F32
    [128]
  • blk.0.attn_norm.weight
    F32
    [128]
  • blk.0.attn_output.bias
    F32
    [128]
  • blk.0.attn_output.weight
    F16
    [128, 128]
  • blk.0.attn_qkv.bias
    F32
    [384]
  • blk.0.attn_qkv.weight
    F16
    [128, 384]
  • blk.0.ffn_down.bias
    F32
    [128]
  • blk.0.ffn_down.weight
    F16
    [512, 128]
  • blk.0.ffn_norm.bias
    F32
    [128]
  • blk.0.ffn_norm.weight
    F32
    [128]
  • blk.0.ffn_up.bias
    F32
    [512]
  • blk.0.ffn_up.weight
    F16
    [128, 512]
    • blk.1
  • blk.1.attn_norm.bias
    F32
    [128]
  • blk.1.attn_norm.weight
    F32
    [128]
  • blk.1.attn_output.bias
    F32
    [128]
  • blk.1.attn_output.weight
    F16
    [128, 128]
  • blk.1.attn_qkv.bias
    F32
    [384]
  • blk.1.attn_qkv.weight
    F16
    [128, 384]
  • blk.1.ffn_down.bias
    F32
    [128]
  • blk.1.ffn_down.weight
    F16
    [512, 128]
  • blk.1.ffn_norm.bias
    F32
    [128]
  • blk.1.ffn_norm.weight
    F32
    [128]
  • blk.1.ffn_up.bias
    F32
    [512]
  • blk.1.ffn_up.weight
    F16
    [128, 512]
    • blk.2
  • blk.2.attn_norm.bias
    F32
    [128]
  • blk.2.attn_norm.weight
    F32
    [128]
  • blk.2.attn_output.bias
    F32
    [128]
  • blk.2.attn_output.weight
    F16
    [128, 128]
  • blk.2.attn_qkv.bias
    F32
    [384]
  • blk.2.attn_qkv.weight
    F16
    [128, 384]
  • blk.2.ffn_down.bias
    F32
    [128]
  • blk.2.ffn_down.weight
    F16
    [512, 128]
  • blk.2.ffn_norm.bias
    F32
    [128]
  • blk.2.ffn_norm.weight
    F32
    [128]
  • blk.2.ffn_up.bias
    F32
    [512]
  • blk.2.ffn_up.weight
    F16
    [128, 512]
    • blk.3
  • blk.3.attn_norm.bias
    F32
    [128]
  • blk.3.attn_norm.weight
    F32
    [128]
  • blk.3.attn_output.bias
    F32
    [128]
  • blk.3.attn_output.weight
    F16
    [128, 128]
  • blk.3.attn_qkv.bias
    F32
    [384]
  • blk.3.attn_qkv.weight
    F16
    [128, 384]
  • blk.3.ffn_down.bias
    F32
    [128]
  • blk.3.ffn_down.weight
    F16
    [512, 128]
  • blk.3.ffn_norm.bias
    F32
    [128]
  • blk.3.ffn_norm.weight
    F32
    [128]
  • blk.3.ffn_up.bias
    F32
    [512]
  • blk.3.ffn_up.weight
    F16
    [128, 512]
  • output_norm.bias
    F32
    [128]
  • position_embd.weight
    F32
    [128, 128]
  • output_norm.weight
    F32
    [128]