Simple WordLevel tokenizer

[iyubondyrev]

Hello!

I’m trying to get a basic word-level tokenizer to work with a smaller version of the Phi3ForCasualML model, which only has 2 layers and 4 heads. The vocabulary is pretty small too, only 382 words.

Here's how I've set up my SentencePiece tokenizer:

model = spm.SentencePieceTrainer.Train(input=files, model_prefix="tokenizer", vocab_size=4, model_type="word", character_coverage=1.0, 
    bos_id=1,
    pad_id=0,
    eos_id=2,
    unk_id=3,
    )

It basically splits sentences into words and assigns numbers to them. But I’m stuck trying to convert this setup into a .gguf format. I have a tokenizer.model from SentencePiece, but it’s not working right now. Here’s what happens when I try to load the model:

llama_model_loader: loaded meta data with 22 key-value pairs and 15 tensors from ggml-model-f32.gguf (version GGUF V3 (latest))
llama_model_loader: Dumping metadata keys/values. Note: KV overrides do not apply in this output.
llama_model_loader: - kv   0:                       general.architecture str              = phi3
llama_model_loader: - kv   1:                               general.name str              = Phi3
llama_model_loader: - kv   2:                        phi3.context_length u32              = 512
llama_model_loader: - kv   3:  phi3.rope.scaling.original_context_length u32              = 512
llama_model_loader: - kv   4:                      phi3.embedding_length u32              = 256
llama_model_loader: - kv   5:                   phi3.feed_forward_length u32              = 1024
llama_model_loader: - kv   6:                           phi3.block_count u32              = 2
llama_model_loader: - kv   7:                  phi3.attention.head_count u32              = 4
llama_model_loader: - kv   8:               phi3.attention.head_count_kv u32              = 4
llama_model_loader: - kv   9:      phi3.attention.layer_norm_rms_epsilon f32              = 0.000010
llama_model_loader: - kv  10:                  phi3.rope.dimension_count u32              = 64
llama_model_loader: - kv  11:                        phi3.rope.freq_base f32              = 10000.000000
llama_model_loader: - kv  12:                          general.file_type u32              = 0
llama_model_loader: - kv  13:                       tokenizer.ggml.model str              = llama
llama_model_loader: - kv  14:                         tokenizer.ggml.pre str              = default
llama_model_loader: - kv  15:                      tokenizer.ggml.tokens arr[str,382]     = ["<pad>", "<s>", "</s>", "<unk>", "�...
llama_model_loader: - kv  16:                      tokenizer.ggml.scores arr[f32,382]     = [0.000000, 0.000000, 0.000000, 0.0000...
llama_model_loader: - kv  17:                  tokenizer.ggml.token_type arr[i32,382]     = [3, 3, 3, 2, 1, 1, 1, 1, 1, 1, 1, 1, ...
llama_model_loader: - kv  18:                tokenizer.ggml.bos_token_id u32              = 1
llama_model_loader: - kv  19:                tokenizer.ggml.eos_token_id u32              = 2
llama_model_loader: - kv  20:            tokenizer.ggml.padding_token_id u32              = 0
llama_model_loader: - kv  21:               general.quantization_version u32              = 2
llama_model_loader: - type  f32:   15 tensors
llm_load_vocab: SPM vocabulary, but newline token not found: _Map_base::at! Using special_pad_id instead.llm_load_vocab: special tokens cache size = 4
llm_load_vocab: token to piece cache size = 0.0028 MB
llama_model_load: error loading model: error loading model vocabulary: _Map_base::at
llama_load_model_from_file: failed to load model
Error: could not load model from file 'ggml-model-f32.gguf'.

Any ideas on how to make this tokenizer work with the transformer model or how to fix the loading issue? Thanks for any help you can offer!