EXLLM-ONI5M
EXLLM-ONI5M is an experimental 0.005377824B-parameter Little Language Model trained to produce short Japanese replies in a deliberately impatient, old-school teacher persona. It is a playful, narrow template-based model—not a general assistant or a reliable source of facts.
EXQ12 inference has been verified on a CASIO EX-word XD-B4800 (DATAPLUS 6). The included GGUF is a separately trained Llama-compatible companion, not a conversion or quantization of the embedded checkpoint.
Highlights
- 0.005377824B-parameter embedded checkpoint with a 128-token context and 868-token vocabulary
- Short kana-oriented prompts and an intentionally gruff fictional persona
- Physical-device inference verified at approximately 0.57 token/s in three recorded XD-B4800 runs
- Separately trained 0.005441184B-parameter GGUF companion for LM Studio and llama.cpp
- Project-authored templates and synthetic combinations; no external pretrained checkpoint
- Apache-2.0
Intended Behavior
The training objective favors a short answer followed by a teacher-like admonition. Some unknown-input examples ask the user to rephrase, but this is learned text behavior—not a rejection mechanism or a guarantee against hallucination.
The persona is fictional. It may produce rude or inaccurate text and has not been evaluated as a safety system. It is intended for embedded inference experiments and entertainment, not factual, professional, or safety-critical use.
Physical EX-word Results
Readback SHA-256 verification, model loading, inference, screenshots, and PERFLOG capture were completed on an XD-B4800. The three recorded runs decoded at approximately 0.57 token/s. One prompt containing kanji did not produce the expected identity response and instead reached the 40-token limit, so these measurements demonstrate execution rather than general response quality.
The machine-readable observations are in benchmarks/xd-b4800-20260930.json.
Usage
LM Studio / llama.cpp
EXLLM-ONI5M-F16.gguf was separately initialized and trained from the ONI5M public training data. It has different weights, tokenizer, architecture details, and runtime context from the EX-word checkpoint.
llama-cli -m EXLLM-ONI5M-F16.gguf -c 512 --single-turn -p "こんにちは"
An RTX 3060 llama.cpp smoke test observed approximately 443 token/s and produced:
話を聞け！あいさつはいい！質問を言え！ 一つずつ確実にやれ！
This is a PC companion smoke-test measurement, not a directly comparable benchmark against the EX-word model. Training details and hashes are in lmstudio/training_manifest.json.
CASIO EX-word
Use exllm-exword v1.3.1 or later with weights/oni5m.q12. The file can be checked before deployment with exllm-model-check.
Training and Provenance
The embedded checkpoint was trained from random initialization for 18 epochs, recording 6,692,919 processed tokens and 4,532,281 loss-bearing tokens. Its data consists of project-authored templates and synthetic combinations. No third-party pretrained checkpoint or real-person quotation corpus was used.
See DATA_PROVENANCE.md, training/RECIPE.md, and release-manifest.json for the release boundary and reproducibility details.
Release Artifacts
| Artifact | Size | SHA-256 | Purpose |
|---|---|---|---|
weights/EXLLM-ONI5M-v1.0.pt | 21,528,721 bytes (20.53 MiB; 0.021528721 GB) | 661857ffc19dbca63f738173d1ef5e6c36e110413028fe88e0e24ad0799d6e7b | FP32 PyTorch checkpoint |
weights/EXLLM-ONI5M-v1.0-int8.bin | 5,443,105 bytes (5.19 MiB; 0.005443105 GB) | 5a6bb0b7c9b15bfcda2cdd50c734b428488e9994a53ef47f9fa49653cf9b9adf | EXLLM8 artifact |
weights/oni5m.q12 | 5,443,105 bytes (5.19 MiB; 0.005443105 GB) | 64c8ce542fc251e26e9e10d756ef0bd33cbdf9e0d7336d1a7048538aca42c5ee | EX-word EXQ12 artifact |
EXLLM-ONI5M-F16.gguf | 10,918,368 bytes (10.41 MiB; 0.010918368 GB) | 6c78fa17860062730bdcfa45835528dc6372b47c0fc28667bd0bc9e4d18483a8 | Separately trained LM Studio / llama.cpp companion |
tokenizer.json | 6,864 bytes | 4c2a583b554e8165936cb3b11727f79d2ede43b73d416a48681f8d38385f5964 | Embedded-model tokenizer |
Limitations
- This is a narrow persona experiment trained on finite templates and synthetic combinations.
- It may answer incorrectly, ignore the intended persona, repeat text, or consume the full generation limit.
- Unknown-input rephrasing is not guaranteed and does not prevent hallucination.
- Kana variation, typos, kanji, long prompts, arithmetic, and open-ended conversation are unreliable.
- The embedded model context is 128 tokens; the GGUF companion's 512-token runtime allocation does not establish 512-token task quality.
- Physical-device and PC companion results belong to different models and numerical runtimes.
- Do not use it for medical, legal, financial, safety-critical, or other professional decisions.
Links
- Shared architecture, conversion, and evaluation project: ToTo-40417/exllm
- EX-word runtime: ToTo-40417/exllm-exword
- Related Japanese-to-English model: ToTo-40417/EXLLM-JPTOEN
- Base EXLLM model: ToTo-40417/EXLLM
日本語
EXLLM-ONI5Mは、短気な昔気質の先生を演じる、5,377,824（0.005377824B）パラメータの遊び寄りのLittle Language Modelです。短いかな中心の入力に対し、短い回答と小言を返すよう学習していますが、汎用会話モデルでも安全機構でもありません。誤答や想定外の応答も発生します。
CASIO XD-B4800（DATAPLUS 6）でEXQ12推論を確認しています。同梱GGUFはEX-word用重みの変換ではなく、公開学習データから別途学習したPC用companionです。詳しい来歴、実行方法、制限事項は上の英語本文と付属資料を参照してください。