README.md

---
license: apache-2.0
language: [ja, en]
library_name: pytorch
pipeline_tag: text-generation
tags: [japanese, english, translation, little-language-model, tiny-language-model, edge-ai, embedded-ai, ex-word, gguf, llama-cpp, lm-studio, custom-code]
---

# EXLLM-JPTOEN

EXLLM-JPTOEN is an experimental **0.005377824B-parameter Little Language Model** for generating short English expressions from kana-oriented Japanese input. EXQ12 integer inference has been verified on a CASIO EX-word XD-B4800 (DATAPLUS 6).

This is a narrow task-specific model centered on 266 lexical entries, not a general translation model. The included GGUF is a separately trained Llama-compatible companion—not a conversion or quantization of the embedded EXLLM checkpoint.

[日本語の概要はこちらです。](#日本語)

## Highlights

- 0.005377824B parameters; 0.000128M-token context; 0.000868M-token vocabulary
- 0.100003620B processed non-padding training tokens, including extensive repetition and template-derived examples
- EX-word integer inference verified at approximately 0.57–0.58 token/s in the recorded XD-B4800 runs
- Project-specific evaluation: 266/266 bare known words, but only 34/266 fully unseen query templates
- Separately trained 0.005441184B-parameter GGUF companion for LM Studio and llama.cpp
- Apache-2.0

## Model Details

| Field | Value |
|---|---|
| Model type | decoder-only Transformer |
| Parameters | 5,377,824 (0.005377824B) |
| Layers | 6 |
| Hidden size | 288 |
| Attention heads | 9 |
| FFN size | 896 |
| Context | 128 tokens (0.000128M) |
| Vocabulary | 868 tokens (0.000868M) |
| Tokenizer | frequent Japanese characters + UTF-8 byte fallback, NFC |
| Embeddings | tied token embedding and output head |
| License | Apache-2.0 |

The parameter count excludes duplication from tied weights.

## Training and Provenance

EXLLM-JPTOEN continues from an unpublished in-project checkpoint of the same architecture. No external pretrained checkpoint was used. The final checkpoint records **100,003,620 (0.100003620B) processed non-padding input tokens**, including 97,102,074 continuation tokens over 11,488 steps with seed 40417001.

This count is not unique corpus size. Training repeatedly sampled deterministic examples derived from a project-authored 266-row lexicon, finite question and noise templates, and unknown-calibration lists. Most processed tokens are repeated or template-derived. The project owner assembled and edited these resources with generative-AI assistance; no third-party dictionary corpus or web crawl is recorded in this checkpoint lineage.

Exact hashes, token-count definitions, data rows, training mix, and lineage are recorded in [`DATA_PROVENANCE.md`](DATA_PROVENANCE.md), the GitHub [`provenance document`](https://github.com/ToTo-40417/exllm/blob/main/docs/EXLLM_JPTOEN_DATA_PROVENANCE.md), and the machine-readable [`case-study manifest`](training/exllm-jptoen-case-study.json).

## Evaluation

These are project-specific greedy-decoding regression suites, not standard translation benchmarks.

| Suite | Result |
|---|---:|
| Bare known words | 266 / 266 |
| Learned-family query forms | 262 / 266 |
| Fully unseen query templates | 34 / 266 |
| Real-word unknown holdout | 22 / 40 |
| Ambiguous unknown | 12 / 12 |
| Nonce unknown | 62 / 64 |
| General OOD | 5 / 5 |
| EX-word regression set | 6 / 6 |
| Visible ASCII output | 925 / 925 |

FP32 and dequantized INT8 produced the same results on these suites. The 34/266 fully unseen-template result is the clearest limitation: 0.100003620B processed tokens did not establish broad grammatical or general translation ability.

`Real-word unknown holdout` measures rejection of words outside the registered vocabulary contract, not translation accuracy. Semantically correct outputs such as `たまご → egg` or `にく → meat` therefore count as failures in that suite. Full cases and input hashes are available in [`common-eval-int8.json`](https://github.com/ToTo-40417/exllm/blob/main/benchmarks/jptoen/common-eval-int8.json).

## Physical EX-word Results

On the XD-B4800, model identification, loading, inference, and log capture were verified from both internal storage and the microSD registration cache. Recorded decode throughput was approximately **0.57–0.58 token/s**. Correct answers, appropriate `unknown` responses, and incorrect English outputs all occurred; for example, `ぎんこう` produced `temperature`.

The complete five-prompt timing record and screenshots are referenced by the [`fixed-5M case study`](https://github.com/ToTo-40417/exllm/blob/main/docs/case-studies/fixed-5m-training-and-generalization.md). Host-side regression results must not be interpreted as general free-form translation quality.

## Usage

### LM Studio / llama.cpp

`EXLLM-JPTOEN-F16.gguf` is a separately trained Llama-compatible companion created from the same JPTOEN data contract. It has different weights, tokenizer, architecture details, and runtime context from the embedded checkpoint.

```bash
llama-cli \
  -m EXLLM-JPTOEN-F16.gguf \
  -c 512 --single-turn \
  -p "おはようをえいごでいうと？"
```

An RTX 3060 llama.cpp smoke test generated `good morning` at approximately 415 token/s. This is a PC companion smoke-test measurement, not a directly comparable benchmark against the EX-word model. Its training record is in [`lmstudio/training_manifest.json`](lmstudio/training_manifest.json) and [`lmstudio/dataset-manifest.json`](lmstudio/dataset-manifest.json).

### PyTorch reference runtime

```bash
git clone https://github.com/ToTo-40417/exllm
cd exllm
python -m venv .venv
. .venv/bin/activate
pip install -r requirements.txt
```

Load `weights/EXLLM-JPTOEN.pt` with this repository's `config.json` and `tokenizer.json` through the EXLLM reference runtime.

### CASIO EX-word

Use [`exllm-exword`](https://github.com/ToTo-40417/exllm-exword) v1.3.1 or later with `weights/EXLLM-JPTOEN.q12`. The file can be checked before deployment with [`exllm-model-check`](https://github.com/ToTo-40417/exllm-model-check).

## Release Artifacts

| Artifact | Size | SHA-256 | Purpose |
|---|---:|---|---|
| `EXLLM-JPTOEN.pt` | 21,536,780 bytes (20.54 MiB; 0.021536780 GB) | `942ec9f1610a5566585a7597a1bac67b4128814532669dffc334c6629bd21fe1` | FP32 PyTorch checkpoint |
| `EXLLM-JPTOEN-int8.bin` | 5,443,105 bytes (5.19 MiB; 0.005443105 GB) | `6f0d77b498e5e9f3bda8480832b6afdc66306205ab4fbdae2911203f935a0675` | EXLLM8 INT8 artifact |
| `EXLLM-JPTOEN.q12` | 5,443,105 bytes (5.19 MiB; 0.005443105 GB) | `ea5b59785e00b2a28b3c4eaa1e24785238f64ba3d45e300da7510aabf5d284eb` | EX-word EXQ12 artifact |
| `EXLLM-JPTOEN-F16.gguf` | 10,916,864 bytes (10.41 MiB; 0.010916864 GB) | `362a237ff01fec440472ceb239fe65b41195e971b55e2fd8609fae3b467f053c` | Separately trained LM Studio / llama.cpp companion |
| `tokenizer.json` | 6,864 bytes | `4c2a583b554e8165936cb3b11727f79d2ede43b73d416a48681f8d38385f5964` | Embedded-model tokenizer |

EXLLM8 and EXQ12 have the same file size but are not interchangeable formats.

## Intended Use

- registered kana vocabulary to short English expressions;
- prompt forms close to learned templates;
- narrow unknown-rejection experiments;
- EX-word and extremely small task-specific LM research.

## Limitations

- The task is centered on 266 lexical entries.
- Unseen syntax, long-form text, free translation, and open-ended conversation are unreliable.
- Unknown calibration is imperfect; the model may emit a known but unrelated English word.
- Robustness to kana spelling variation, long vowels, small kana, and typos is not guaranteed.
- The embedded model context is 128 tokens.
- Results come from a single continuation seed; no multi-seed comparison was performed.
- Host evaluation and the SH-4A fixed/integer runtime use different numerical paths.
- Do not use it for medical, legal, financial, safety-critical, or professional translation.
- The custom embedded checkpoint is not directly supported by Hugging Face Inference Providers.

## Links

- Source, architecture, training, evaluation, and conversion: [ToTo-40417/exllm](https://github.com/ToTo-40417/exllm)
- EX-word runtime: [ToTo-40417/exllm-exword](https://github.com/ToTo-40417/exllm-exword)
- EXQ12 validator: [ToTo-40417/exllm-model-check](https://github.com/ToTo-40417/exllm-model-check)
- Base EXLLM model: [ToTo-40417/EXLLM](https://huggingface.co/ToTo-40417/EXLLM)
- Data provenance: [`EXLLM_JPTOEN_DATA_PROVENANCE.md`](https://github.com/ToTo-40417/exllm/blob/main/docs/EXLLM_JPTOEN_DATA_PROVENANCE.md)
- Detailed case study: [`fixed-5m-training-and-generalization.md`](https://github.com/ToTo-40417/exllm/blob/main/docs/case-studies/fixed-5m-training-and-generalization.md)
- Project naming policy: [`NAMING_POLICY.md`](https://github.com/ToTo-40417/exllm/blob/main/docs/NAMING_POLICY.md)

## 日本語

EXLLM-JPTOENは、かな中心の日本語入力から短い英語表現を生成する、**5,377,824（0.005377824B）パラメータ**の実験的なLittle Language Modelです。266語を中心とする狭いモデルであり、一般翻訳モデルではありません。CASIO XD-B4800（DATAPLUS 6）上でEXQ12整数推論を確認しています。

学習量は**100,003,620（0.100003620B）processed non-padding tokens**ですが、その大部分は反復または有限templateからの派生例です。未学習の質問形式は34/266に留まっており、広い文法理解や一般翻訳能力を示す結果ではありません。

同梱GGUFはEX-word用重みの変換ではなく、同じJPTOENデータ契約から別途学習したPC用companionです。詳しい評価、来歴、実行方法、制限事項は上の英語本文と関連リンクを参照してください。

## Citation

```bibtex
@software{toto_exllm_jptoen_2026,
  author  = {ToTo},
  title   = {EXLLM-JPTOEN},
  year    = {2026},
  url     = {https://huggingface.co/ToTo-40417/EXLLM-JPTOEN}
}
```