---
library_name: transformers
license: mit
datasets:
- HuggingFaceFW/fineweb
language:
- en
base_model:
- sign/utf8-lm-tiny
---

# UTF16-LM-tiny

This model is a fine-tuned version of [sign/utf8-lm-tiny](https://huggingface.co/sign/utf8-lm-tiny) on the HuggingFaceFW/fineweb dataset.

Using [this](https://github.com/sign/utf8-tokenizer/blob/main/experiments/language-modelling/run_clm.py) training script, from [utf8-tokenizer](https://github.com/sign/utf8-tokenizer/tree/main).

Unlike the base model, where we train directly on UTF-8 bytes, here we train on UTF-16 code units. Each Unicode character is represented as **one or two UTF-16 code units** (surrogate pairs for non-BMP characters). Each code unit is decomposed into two bytes, which are encoded independently and then concatenated.

| Character | UTF-8              | UTF-16                  | UTF-16 Decomposed (bytes) |
| --------- | ------------------ | ----------------------- | ------------------------- |
| A         | `\x41`             | `U+0041`                | `[0, 65]`                 |
| é         | `\xC3\xA9`         | `U+00E9`                | `[0, 233]`                |
| €         | `\xE2\x82\xAC`     | `U+20AC`                | `[32, 172]`               |
| 😀        | `\xF0\x9F\x98\x80` | `U+D83D U+DE00`         | `[216, 61] , [222, 0]`    |

This replaces UTF-8’s **1–4 byte** variable-length encoding with a **1–2 code-unit** representation. While still variable-width, UTF-16 substantially reduces sequence-length variance and upper-bounds expansion for complex scripts and emoji-heavy text compared to UTF-8.  
By contrast, the [utf32-lm-tiny](https://huggingface.co/sign/utf32-lm-tiny) model uses four code units per character (fixed-width), yielding the simplest mapping from Unicode scalars to token sequences at the cost of longer overall sequences.


## Usage

```python
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/utf16-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

```shell
python run_clm.py \
  --use_bit_embeddings False \
  --encoding utf16 \
  --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