About the developer — One person (FiShota) building a Japanese LM stack from scratch on a single RTX 3090. HinoMoto = from-scratch JP LM family, Yamato = legal/admin SFT specialist, HinoMoto-Bench-ja = cultural-axis benchmark. Honest research: every release links to a write-up that includes the negative results. GitHub · Bench · X

Quick facts

Parameters: 100M (12 layer, d_model=512, vocab=9506)
Tokenizer: Byte-BPE 9506 tokens (balanced corpus)
Training: 20000 steps, RTX 3090 24GB, fp32, ~28 min
Schedule: WSD (Warmup-Stable-Decay), warmup=400, stable=80%, decay=20%
Regularizer: z-loss coef=1e-4
EMA: decay=0.999 (CPU shadow)

PPL (51 / 200 / 789 prompts)

Eval set	PPL
51 short prompts	15.94
200 mixed prompts	32.08
789 broad prompts	48.51

Honest comparison with v12 (WSD seed=2) and cosine baselines

3-seed × 2-schedule × 3 eval-set 計測:

Schedule	seed=0	seed=1	seed=2	mean (51)	std (51)	mean (789)	std (789)
Cosine	v7 (17.72)	v8 (17.09)	v14 (16.58)	17.13	0.57	48.03	0.18
WSD+z-loss	v15 (15.94)	v11/v16 (16.50)	v12 (16.35)	16.26	0.29	48.76	0.45

paired t-test (cosine vs WSD): p=0.148 (51), p=0.484 (200), p=0.534 (789) = 全 eval で有意差なし.

→ WSD ≈ cosine at 100M JP PPL (no statistically significant difference).

Update — 2026-05-09 — Phase 1 (350M scale-up) progress

The HinoMoto pipeline has scaled up to 350M params with the same architecture family (d_model=1024, 24 layers, vocab=9506).

Run	Params	Schedule	Steps	Final ppl	Notes
100M v15 (this card)	100M (43M excl embed)	WSD+z-loss+EMA	20,000	15.94 (51-prompt)	this card
350M smoke	318M	WSD+z-loss+EMA bf16	5,000	9.53 (training)	architecture verified
350M Phase 1 full	318M	WSD+z-loss+EMA bf16	50,000	(in progress)	RTX 3090 single GPU, ~22h

Architecture scales cleanly from 100M → 350M with the same recipe. Next: HinoMoto-1B (Phase 2) after corpus expansion (78MB → 5GB+ planned).

→ Detailed retro: see HinoMoto 開発ノート #10 (post Phase 1 完了後).

How to use

python

import torch
from hinomoto.model.hinomoto_model import HinoMotoModel, HinoMotoConfig
from hinomoto.tokenizer.byte_bpe import ByteBPETokenizer

ck = torch.load("ckpt_ema_step_020000_final.pt", map_location="cpu", weights_only=False)
cfg = HinoMotoConfig(**ck["config"])
model = HinoMotoModel(cfg)
model.load_state_dict(ck["shadow"], strict=False)
model.lm_head.weight = model.tok_embed.weight  # tied
model.eval()

tok = ByteBPETokenizer.load("tokenizer.json")
ids = tok.encode("日本語の文章モデルを作るには")
# ... your inference loop

Training command

bash

python -m hinomoto.train.train_lm \
  --tokenizer tokenizer_v3_32k_clean.json \
  --corpus all_v6_balanced.txt \
  --output-dir artifacts/smoke_100m_v15_wsd_zloss_ema \
  --config configs/main_run_100m_v3.json \
  --max-steps 20000 --warmup-steps 400 \
  --batch-size 2 --grad-accum 4 --seq-len 512 \
  --lr 3e-4 --device cuda \
  --dtype fp32 --seed 0 \
  --ema-decay 0.999 \
  --lr-schedule wsd --wsd-decay-frac 0.2 \
  --z-loss-coef 1e-4 \
  --spike-detect

Reproducibility

The HinoMoto pipeline is fp32 bit-deterministic in our environment (Windows 11 + WSL2 Ubuntu 24.04 + RTX 3090). Same --seed → same weights, even across days/runs (verified: v11 = v16 with seed=1).

License

Apache-2.0

Citation

Coming soon (arXiv WIP, see paper_drafts/hinomoto_arxiv_outline.md).

All FiShota models (full family map)

From-scratch JP LM (HinoMoto, Apache-2.0)

Repo	Scale	Stage	Status
hinomoto-100m-v15-wsd-zloss-ema	100M	research baseline (this card)	✅ stable
hinomoto-100m-v12-wsd-zloss-seed2	100M	sister seed=2 (reproducibility)	✅ stable

Sarashina2.2-3B-instruct + cultural-axis SFT (MIT)

Repo	Recipe	Notes
sarashina2.2-3b-sft-v3-Q4_K_M-gguf	4-axis SFT, 4 quants (Q3/Q4/Q5/Q6)	family / keigo / silence / atmosphere
sarashina2.2-3b-sft-v4-gguf	sft_v3 + 11% NLI replay	catastrophic-forgetting repair
sarashina2.2-3b-sft-v4-dpo-gguf	sft_v4 + DPO 108 pairs	preference tuning

Yamato — Japanese legal / administrative specialist (MIT, LoRA r-ablation)

Repo	LoRA r	Bench %	Notes
yamato-3b-v1-legal-gguf	16	46.7%	baseline
yamato-3b-v2-legal-gguf	16	—	2nd pass
yamato-3b-v3-r32-legal-gguf	32	48.9%	+2.2
yamato-3b-v4-r64-legal-gguf	64	54.3%	+5.4
yamato-3b-v5-r128-legal-gguf ⭐	128	57.6%	best — see dev-note #9

Companion repos (GitHub)

hinomoto-bench-ja — cultural-axis benchmark, CC BY 4.0, 418+ items
hinomoto-model — training pipeline (this card's recipe)
hinomoto-mojo — Mojo SIMD/parallel kernels for inference
rtk — Rust corpus-cleaning tool

📊 350M Phase 1 training curve (live)

↑ HinoMoto-350M Phase 1 (in-progress). Updated periodically. Left: perplexity (log scale). Right: loss (linear). See HinoMoto 開発ノート #10 (post-completion retro).

FiShota/hinomoto-100m-v15-wsd-zloss-ema

HinoMoto-100M v15 (WSD + z-loss + EMA, seed=0)