evanollama/ tiny-stories-50m-release:latest

38 3 weeks ago

I trained a 51M parameter GPT-style model on an RTX 3080 10GB using TinyStories. It can generate coherent children’s-story style text. Includes HF Transformers weights and GGUF/Ollama export. Followed https://arxiv.org/abs/2305.07759

... /
model
e86812b5905f · 105MB
    Metadata
  • general.architecture
    gpt2
  • general.file_type
    F16
  • gpt2.attention.head_count
    8
  • gpt2.attention.layer_norm_epsilon
    1e-05
  • gpt2.block_count
    8
  • gpt2.context_length
    512
  • gpt2.embedding_length
    512
  • gpt2.feed_forward_length
    2048
  • tokenizer.ggml.bos_token_id
    50256
  • tokenizer.ggml.eos_token_id
    50256
  • tokenizer.ggml.merges
    [Ġ t, Ġ a, h e, i n, r e, ...]
  • tokenizer.ggml.model
    gpt2
  • tokenizer.ggml.padding_token_id
    50256
  • tokenizer.ggml.pre
    gpt-2
  • tokenizer.ggml.token_type
    [1, 1, 1, 1, 1, ...]
  • tokenizer.ggml.tokens
    [!, ", #, $, %, ...]
  • tokenizer.ggml.unknown_token_id
    50256
    • Tensor
  • token_embd.weight
    F16
    [512, 50304]
    • blk.0
  • blk.0.attn_norm.bias
    F32
    [512]
  • blk.0.attn_norm.weight
    F32
    [512]
  • blk.0.attn_output.bias
    F32
    [512]
  • blk.0.attn_output.weight
    F16
    [512, 512]
  • blk.0.attn_qkv.bias
    F32
    [1536]
  • blk.0.attn_qkv.weight
    F16
    [512, 1536]
  • blk.0.ffn_down.bias
    F32
    [512]
  • blk.0.ffn_down.weight
    F16
    [2048, 512]
  • blk.0.ffn_norm.bias
    F32
    [512]
  • blk.0.ffn_norm.weight
    F32
    [512]
  • blk.0.ffn_up.bias
    F32
    [2048]
  • blk.0.ffn_up.weight
    F16
    [512, 2048]
    • blk.1
  • blk.1.attn_norm.bias
    F32
    [512]
  • blk.1.attn_norm.weight
    F32
    [512]
  • blk.1.attn_output.bias
    F32
    [512]
  • blk.1.attn_output.weight
    F16
    [512, 512]
  • blk.1.attn_qkv.bias
    F32
    [1536]
  • blk.1.attn_qkv.weight
    F16
    [512, 1536]
  • blk.1.ffn_down.bias
    F32
    [512]
  • blk.1.ffn_down.weight
    F16
    [2048, 512]
  • blk.1.ffn_norm.bias
    F32
    [512]
  • blk.1.ffn_norm.weight
    F32
    [512]
  • blk.1.ffn_up.bias
    F32
    [2048]
  • blk.1.ffn_up.weight
    F16
    [512, 2048]
    • blk.2
  • blk.2.attn_norm.bias
    F32
    [512]
  • blk.2.attn_norm.weight
    F32
    [512]
  • blk.2.attn_output.bias
    F32
    [512]
  • blk.2.attn_output.weight
    F16
    [512, 512]
  • blk.2.attn_qkv.bias
    F32
    [1536]
  • blk.2.attn_qkv.weight
    F16
    [512, 1536]
  • blk.2.ffn_down.bias
    F32
    [512]
  • blk.2.ffn_down.weight
    F16
    [2048, 512]
  • blk.2.ffn_norm.bias
    F32
    [512]
  • blk.2.ffn_norm.weight
    F32
    [512]
  • blk.2.ffn_up.bias
    F32
    [2048]
  • blk.2.ffn_up.weight
    F16
    [512, 2048]
    • blk.3
  • blk.3.attn_norm.bias
    F32
    [512]
  • blk.3.attn_norm.weight
    F32
    [512]
  • blk.3.attn_output.bias
    F32
    [512]
  • blk.3.attn_output.weight
    F16
    [512, 512]
  • blk.3.attn_qkv.bias
    F32
    [1536]
  • blk.3.attn_qkv.weight
    F16
    [512, 1536]
  • blk.3.ffn_down.bias
    F32
    [512]
  • blk.3.ffn_down.weight
    F16
    [2048, 512]
  • blk.3.ffn_norm.bias
    F32
    [512]
  • blk.3.ffn_norm.weight
    F32
    [512]
  • blk.3.ffn_up.bias
    F32
    [2048]
  • blk.3.ffn_up.weight
    F16
    [512, 2048]
    • blk.4
  • blk.4.attn_norm.bias
    F32
    [512]
  • blk.4.attn_norm.weight
    F32
    [512]
  • blk.4.attn_output.bias
    F32
    [512]
  • blk.4.attn_output.weight
    F16
    [512, 512]
  • blk.4.attn_qkv.bias
    F32
    [1536]
  • blk.4.attn_qkv.weight
    F16
    [512, 1536]
  • blk.4.ffn_down.bias
    F32
    [512]
  • blk.4.ffn_down.weight
    F16
    [2048, 512]
  • blk.4.ffn_norm.bias
    F32
    [512]
  • blk.4.ffn_norm.weight
    F32
    [512]
  • blk.4.ffn_up.bias
    F32
    [2048]
  • blk.4.ffn_up.weight
    F16
    [512, 2048]
    • blk.5
  • blk.5.attn_norm.bias
    F32
    [512]
  • blk.5.attn_norm.weight
    F32
    [512]
  • blk.5.attn_output.bias
    F32
    [512]
  • blk.5.attn_output.weight
    F16
    [512, 512]
  • blk.5.attn_qkv.bias
    F32
    [1536]
  • blk.5.attn_qkv.weight
    F16
    [512, 1536]
  • blk.5.ffn_down.bias
    F32
    [512]
  • blk.5.ffn_down.weight
    F16
    [2048, 512]
  • blk.5.ffn_norm.bias
    F32
    [512]
  • blk.5.ffn_norm.weight
    F32
    [512]
  • blk.5.ffn_up.bias
    F32
    [2048]
  • blk.5.ffn_up.weight
    F16
    [512, 2048]
    • blk.6
  • blk.6.attn_norm.bias
    F32
    [512]
  • blk.6.attn_norm.weight
    F32
    [512]
  • blk.6.attn_output.bias
    F32
    [512]
  • blk.6.attn_output.weight
    F16
    [512, 512]
  • blk.6.attn_qkv.bias
    F32
    [1536]
  • blk.6.attn_qkv.weight
    F16
    [512, 1536]
  • blk.6.ffn_down.bias
    F32
    [512]
  • blk.6.ffn_down.weight
    F16
    [2048, 512]
  • blk.6.ffn_norm.bias
    F32
    [512]
  • blk.6.ffn_norm.weight
    F32
    [512]
  • blk.6.ffn_up.bias
    F32
    [2048]
  • blk.6.ffn_up.weight
    F16
    [512, 2048]
    • blk.7
  • blk.7.attn_norm.bias
    F32
    [512]
  • blk.7.attn_norm.weight
    F32
    [512]
  • blk.7.attn_output.bias
    F32
    [512]
  • blk.7.attn_output.weight
    F16
    [512, 512]
  • blk.7.attn_qkv.bias
    F32
    [1536]
  • blk.7.attn_qkv.weight
    F16
    [512, 1536]
  • blk.7.ffn_down.bias
    F32
    [512]
  • blk.7.ffn_down.weight
    F16
    [2048, 512]
  • blk.7.ffn_norm.bias
    F32
    [512]
  • blk.7.ffn_norm.weight
    F32
    [512]
  • blk.7.ffn_up.bias
    F32
    [2048]
  • blk.7.ffn_up.weight
    F16
    [512, 2048]
  • output_norm.bias
    F32
    [512]
  • position_embd.weight
    F32
    [512, 512]
  • output_norm.weight
    F32
    [512]