README.md

---
license: apache-2.0
base_model:
- Qwen/Qwen3.5-4B
library_name: mlx
pipeline_tag: text-generation
language:
- en
- zh
- ja
- hi
- ar
tags:
- mlx
- transliteration
- pinyin
- romaji
- hinglish
- arabizi
- qwen3.5
---

> [!IMPORTANT]
> **Newer Generation Available (`2.0`)**: See **[`UnimeType/Transliteration-2.0-4B`](https://huggingface.co/UnimeType/Transliteration-2.0-4B)** for the latest **UnimeType Transliteration 2.0 4B** release (**87.4%** 4-language strict accuracy, `+31.8%` over v1.0, dynamic user glossary `lexicon`, and zero protected-span violations).

# UnimeType Transliteration 4B

Type with familiar Latin characters. Get the writing system your message needs.

UnimeType Transliteration 4B is a focused language conversion model for everyday, mixed-language writing. It converts Pinyin, Romaji, Romanized Hindi, Hinglish, and Arabizi, with English words, product names, links, code, numbers, punctuation, and emoji intended to stay in place. Review outputs before using them; preservation is not guaranteed.

It was created for the Convert experience from [UnimeType](https://unimetype.com/).

## Get the UnimeType keyboard app

Write with Latin letters, then choose **Convert** to change romanized text into your selected writing system. UnimeType also includes **Polish** and **Explain**.

**[Download the UnimeType app →](https://unimetype.com/download?utm_source=huggingface&utm_medium=referral&utm_campaign=transliteration_4b&utm_content=readme_top)**

Choose your platform: iPhone & iPad, Android, Mac, or Windows (beta).

To run this model yourself, see **Download model weights** below.

## Download model weights

| Format | Download |
| --- | --- |
| MLX 6-bit | [This repository](https://huggingface.co/UnimeType/Transliteration-4B/tree/main) |
| Standard floating-point Safetensors | [Safetensors repository](https://huggingface.co/UnimeType/Transliteration-4B-Safetensors/tree/main) |
| GGUF Q8_0 | [GGUF repository](https://huggingface.co/UnimeType/Transliteration-4B-GGUF/tree/main) |

The new exports share the retained fine-tuned source described below. Their model cards document a known Arabic trailing-newline error and runtime-specific validation. This repository's MLX weights remain unchanged.

## Four writing systems, one familiar input

| Input style | Example input | Model output |
| --- | --- | --- |
| Chinese Pinyin | `jintian yao finish README` | `今天要 finish README` |
| Japanese Romaji | `ashita Tokyo de meeting ga arimasu` | `明日 Tokyo で meeting があります` |
| Romanized Hindi / Hinglish | `aaj meeting hai` | `आज meeting है` |
| Arabic Arabizi | `fee meeting bokra` | `في meeting بكرة` |

These examples are direct outputs from the published 6-bit model in LM Studio.

## Made for real messages

UnimeType Transliteration 4B is designed for text that does not fit into a traditional transliteration box:

- Mixed-language sentences that combine romanized text with English terms
- Short chat messages and conversational phrases
- Product names, technical vocabulary, URLs, email addresses, and code spans
- Punctuation, line breaks, Markdown, numbers, and emoji that should remain intact
- Text that is already in the target writing system and should not be changed

The model returns the converted text directly, ready to review and use.

## Supported conversion modes

| Target | Latin input | Output |
| --- | --- | --- |
| Simplified Chinese | Pinyin and natural unmarked Pinyin | Chinese characters |
| Japanese | Romaji | Kanji, Hiragana, and Katakana |
| Hindi | Romanized Hindi and Hinglish | Devanagari with preserved English |
| Arabic | Arabizi and conversational romanized Arabic | Arabic script with preserved English |

## Run locally

This repository contains an MLX model for Apple silicon Macs. The model can be downloaded from Hugging Face and loaded with MLX-LM or imported into an MLX-compatible local runtime such as LM Studio.

Recommended generation settings:

| Setting | Value |
| --- | --- |
| Thinking / reasoning | Off |
| Temperature | `0` |
| Maximum output | `128` tokens for short messages |
| Context used for validation | `2048` tokens |

Select one target language from `chinese`, `japanese`, `hindi`, or `arabic`, then provide the original text. The model is built for conversion, not translation, rewriting, polishing, or explanation.

## Model details

| Property | Value |
| --- | --- |
| Base model | [Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) |
| Parameters | 4B class |
| Format | MLX |
| Quantization | 6-bit affine, group size 64 |
| Model file size | 3.42 GB |
| Primary task | Context-aware transliteration and script conversion |
| License | Apache 2.0 |

## Historical standalone evaluation

The published model was tested with exact final-text matching. A result passes only when the complete output matches an accepted answer, including preserved English, punctuation, code, and spacing.

| Runtime | Everyday conversion suite | Reserved multilingual suite |
| --- | ---: | ---: |
| Direct MLX | 98 / 127 | 57 / 104 |
| LM Studio | 94 / 127 | 57 / 104 |

The reserved multilingual score measures exact agreement with its reference responses. Exact-match scoring is strict: a valid alternative spelling or wording can still count as a mismatch. Review ambiguous names, dialect terms, and very short fragments before sending them.

## Product-pipeline validation

On September 6, 2026, the unchanged published MLX weights were tested with UnimeType's production Convert prompts and evaluator, using LM Studio on an Apple M5 Pro MacBook Pro with 48 GB memory and a 16,384-token context setting.

| Check | Result |
| --- | ---: |
| Fixed conversion suite, exact accepted output | 97 / 127 |
| Conversational-context regression suite, valid-reference exact agreement | 12 / 71 |
| Context-suite structure and protected-literal count checks | 74 / 76 |
| Protected-literal count and order, cases containing protected literals | 31 / 32 |
| Latin-word preservation in eligible fixed-suite cases | 43 / 48 |
| URL preservation in eligible context cases | 16 / 16 |
| Code-span preservation in eligible context cases | 16 / 16 |
| Numeric-literal preservation in eligible context cases | 16 / 16 |
| Punctuation-sequence preservation in eligible context cases | 15 / 16 |

These are product-pipeline results, including input protection, output validation, and 20 local-route cases in the fixed suite. They are not raw-model accuracy. Five structurally invalid reference answers are excluded from context agreement, but all 76 cases remain in the structure check. Context references are model-generated and have not been reviewed by native speakers. These suites have been used during development and are not an independent blind benchmark.

Preservation counts measure exact literal counts and order within eligible cases, not general language accuracy. The Latin-word check covers case-sensitive words present in the input and every accepted reference, excluding URL, email, and code spans. Context preservation checks cover explicitly marked literals; categories overlap.

Known limitations include ambiguous short inputs, Arabic dialect and spelling variation, and context-related formatting errors. Short English fragments can be converted instead of preserved. A Japanese regression case also failed to preserve punctuation and an emoji. NVIDIA CUDA, AMD ROCm, iOS, and Android inference have not been validated for this release. The historical standalone scores above use a different evaluation setup and should not be treated as interchangeable with this product-pipeline run.

## Release provenance and files

This repository contains the retained MLX 6-bit release. The model was fine-tuned with LoRA on a quantized 4-bit Qwen3.5-4B base, fused into a floating-point checkpoint, and then quantized to 6-bit MLX. Floating-point storage does not recover the original base model's pre-quantization precision.

The retained floating-point checkpoint was audited against this release: its 426 tensors reproduce all 924 published tensors under 6-bit affine quantization. It is not a dequantization of the published 6-bit weights. See [PROVENANCE.json](PROVENANCE.json) for source hashes and [SHA256SUMS](SHA256SUMS) for public-file checksums.

**训练数据集未公开。** The training dataset is not publicly released. This repository does not include training data, data-generation scripts, training code, or training logs.

The base model uses [Apache 2.0](https://huggingface.co/Qwen/Qwen3.5-4B/blob/main/LICENSE). A copy is included in [LICENSE](LICENSE). Safetensors and GGUF downloads are in the separate repositories linked above.

## UnimeType keyboard app

Use Convert, Polish, and Explain from the UnimeType keyboard.

**[Choose your platform and download →](https://unimetype.com/download?utm_source=huggingface&utm_medium=referral&utm_campaign=transliteration_4b&utm_content=readme_bottom)**