UTF32-LM-tiny

This model is a fine-tuned version of sign/utf8-lm-tiny on the HuggingFaceFW/fineweb dataset.

Using this training script, from utf8-tokenizer.

Unlike the base model, where we train directly on UTF-8 bytes - here, we train on characters (UTF-32 blocks), where each character is decomposed into a fixed, four "bytes" which are encoded independently and then concatenated.

Character	UTF-8	UTF-32	UTF-32 Decomposed (bytes)
A	\x41	U+00000041	[0, 0, 0, 65]
é	\xC3\xA9	U+000000E9	[0, 0, 0, 233]
€	\xE2\x82\xAC	U+000020AC	[0, 0, 32, 172]
😀	\xF0\x9F\x98\x80	U+0001F600	[0, 1, 246, 0]

This is effectively switching from variable-width canonical UTF-8 byte sequences, to fixed size groups, making training/inference up to 4x more efficient for complex scripts.

Usage

from transformers import AutoModelForCausalLM, LogitsProcessorList
import torch
from utf8_tokenizer.logits_processor import UTF8ValidationLogitsProcessor
from utf8_tokenizer.char_causal_lm import CharacterCausalLMWrapper

from utf8_tokenizer import UTF8Tokenizer

model_id = "sign/utf32-lm-tiny"

tokenizer = UTF8Tokenizer()
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "My name is"

inputs = tokenizer([prompt], return_tensors="pt",
                   padding=True,
                   add_special_tokens=True)
# We need to remove the EOS token
inputs["input_ids"] = inputs["input_ids"][:, :-1]
inputs["attention_mask"] = inputs["attention_mask"][:, :-1]


with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=256,
    )

print(tokenizer.decode(out[0], skip_special_tokens=False))

Training procedure

python run_clm.py \
  --use_bit_embeddings True \
  --encoding utf32 \
  --output_dir ./output-tiny-lm-fineweb-groups \
  --dataset_name HuggingFaceFW/fineweb \
  --streaming True \
  --dataloader_num_workers 1 \
  --dataloader_prefetch_factor 4 \
  --dataloader_pin_memory True \
  --dataloader_persistent_workers True \
  --do_train True \
  --save_strategy steps \
  --max_steps 100000 \
  --save_steps 1000 \
  --save_total_limit 1 \
  --logging_steps 100 \
  --logging_strategy steps \
  --model_name_or_path sbintuitions/tiny-lm \
  --per_device_train_batch_size 256 \
  --block_size 256 \
  --optim adamw_torch_fused \
  --learning_rate 3e-4 \
  --lr_scheduler_type cosine \
  --warmup_ratio 0.01 \
  --weight_decay 0.1 \
  --adam_beta1 0.9 \
  --adam_beta2 0.95 \
  --max_grad_norm 1.0 \
  --gradient_checkpointing True \
  --bf16 True \
  --seed 42 \
  --report_to wandb \
  --include_num_input_tokens_seen True

Framework versions

• Transformers 4.57.3
• Pytorch 2.9.1+cu130
• Datasets 4.4.1
• Tokenizers 0.22.1