ChiTako/unofficial-photon-repro-ja-250m-v4

Unofficial. This is an independent, hobby reimplementation of the PHOTON architecture described in arXiv:2512.20687, trained from scratch by a private individual. It is not released by, affiliated with, or endorsed by the paper's authors or any organisation, and no weights, data or code from any official PHOTON release were used. Treat it as a research artefact, not a product.

Correction to the perplexity numbers (2026-07-28)

Earlier revisions of this card reported Wikipedia perplexity measured by streaming wikimedia/wikipedia from record 0 — which is where prepare_data.py also starts building the training shards. Those numbers were measured on training data and are withdrawn.

The table below uses articles past the ones training consumed (--ppl-skip-ja 700000 --ppl-skip-en 400000).

Held-out Wikipedia, 409,600 tokens each, identical protocol for all runs:

model	data	ja-wiki	en-wiki
unofficial-photon-repro-ja-250m-v1	Wikipedia only, 200M tok	60.32	65.91
unofficial-photon-repro-ja-250m-v3	Wikipedia only, 200M tok	57.12	62.17
unofficial-photon-repro-ja-250m-v4	Wikipedia only, 200M tok	52.11	59.52
unofficial-photon-repro-ja-250m-v3-900m	full mixture, 900M tok	67.60	69.37

Do not rank these models by the Wikipedia column. v1-v4 trained on Wikipedia; v3-900m did not (5% of its mixture). On 青空文庫, which none of them trained on, v3-900m scores 237.10 against v4's 365.49 — the ranking reverses. See docs/findings.md sections 1b-1d in the repository.

PHOTON-JP small v4

A Japanese language model on the PHOTON hierarchical autoregressive architecture (arXiv:2512.20687), with every transformer stack replaced by a fine-grained Mixture-of-Experts using auxiliary-loss-free load balancing.

Tokens are folded into multi-resolution latent units instead of being scanned one at a time:

tokens   t1 t2 t3 t4 | t5 t6 t7 t8 | ...      <- level-1 decoder, every token
           \  \  /  /
level-1      u1      |     u2      | ...      <- once per 4 tokens
              \_____/____________/
level-2            m1                          <- once per 16 tokens

A level-l encoder runs once per C<=l tokens, so its cost is amortised. That is the whole point: capacity sits at the top of the hierarchy where it is cheapest per token.

Size

total parameters	0.252 B
active per token	0.100 B
FLOP-equivalent dense size	0.069 B
forward FLOPs / token	0.130 GFLOP
hierarchy	L=2, C≤L=16 tokens per top-level unit
context	1024 tokens
vocabulary	99,584 (llm-jp-tokenizer v3)
KV cache, HierGen	0.95 KiB/token
KV cache, RecGen	0.20 KiB/token

Training

tokens seen	0.20 B
steps	1,520
tokens / parameter	0.77
hardware	1x NVIDIA GB10 (DGX Spark), 5 h
throughput	11.1K tokens/s
optimiser	Muon (2-D weights) + AdamW (embeddings, norms, router)
schedule	WSD, 1-sqrt cooldown
final train CE	3.9335 (ppl 51.09)

Evaluation

benchmark	result
ppl/wiki_ja	52.1143
ppl/wiki_en	59.5171

Data

Japanese-majority mixture: fineweb-2-edu-japanese, FineWeb-2 ja, Japanese Wikipedia, Zyda-2, FinePDFs-Edu, FineWeb-Edu and SwallowMath-v2, tokenised with llm-jp-tokenizer v3. The last 15% of training switches to a high-quality math / code / Japanese-edu blend (Nemotron-3, OLMo 3 and llm-jp-4 all do this).

Honest limitations

This model is under-trained by design of the budget, not by accident. At 0.77 tokens per parameter it is far below the ~20 that Chinchilla-optimal training implies, and further still below what an inference-efficient model would normally get. It produces fluent Japanese surface form -- correct particles, natural kana/kanji mixing, sentence-final forms -- while being largely incoherent semantically. Treat it as a demonstration that the architecture trains, not as a usable assistant.

Samples

九星陵から西に向かって高句麗の帝都を立てびわこじている。

首都は西西方、東は東南端部の西北西に位置し、東に北の東で


温泉街 
峡谷を形成する広大な区域に立地し、山の麓は檜山山脈にそびえている。近くに能勢寺があり、


生体工学
人間工学

五代桂桂春画に代表される、10歳が高かった「望林」の絵図を書いた『古志曲屏風』を描いた本郷台の主たる人物である。

蘇雲河江に面した林田


石英の雫石
石石に含まれる石英を塩基の成分アル

Generation protocols

• HierGen keeps encoder state at every level. Exact: it reproduces the training-time distribution (verified to 2e-4 in the test suite).
• RecGen keeps only the top-level KV cache and feeds the decoder cascade its own reconstructions, cutting cache by 4.8x. Exact when recursive consistency holds, which is what L_token + alpha * L_rec optimises for.

Usage

from photon_jp.model.config import PhotonConfig
from photon_jp.model.photon import PhotonForCausalLM
from photon_jp.model.loading import load_state_dict_compat
from photon_jp.infer.generate import PhotonGenerator, GenerationConfig
from safetensors.torch import load_file

cfg = PhotonConfig.load("model_config.json")
model = PhotonForCausalLM(cfg)
load_state_dict_compat(model, load_file("model.safetensors"))

gen = PhotonGenerator(model.cuda().eval(), "cuda")
out = gen.generate(input_ids, GenerationConfig(mode="recgen", max_new_tokens=256))

Accounting

==============================================================================
PHOTON-JP parameter report
==============================================================================
vocab=99,584  D0=512  L=2  C_<=L=16  ctx=1024

stack                    total        active   amort     amort.act
------------------------------------------------------------------------------
L1.encoder             102.7M        12.3M       4         3.1M
L1.decoder              62.8M         7.9M       1         7.9M
L2.encoder              19.9M        19.9M      16         1.2M
L2.decoder               8.6M         4.5M       4         1.1M
------------------------------------------------------------------------------
embedding               51.0M
lm_head                  0.0M
chunk/convert            4.2M
mtp                      2.7M
==============================================================================
TOTAL          : 0.252 B   (0.201 B non-emb)
ACTIVE / token : 0.100 B   (0.049 B non-emb)
AMORTISED      : 0.069 B  (FLOP-equivalent dense size)
sparsity       : 2.52x

fwd FLOPs/token @ ctx=1024: 0.13 GFLOP (matmul 0.13, attn 0.001)
KV cache mode    : MLA latent (weight-absorbed)
KV cache  HierGen: 0.945 KiB/token   (other mode: 2.312)
KV cache  RecGen : 0.195 KiB/token   (other mode: 1.562)
  -> 0.9 MiB for a full 1024-token context (HierGen, per sequence)
==============================================================================

Citation

@article{ichikawa2025photon,
  title  = {PHOTON: Hierarchical Autoregressive Modeling for Lightspeed and
            Memory-Efficient Language Generation},
  author = {Ichikawa, Yuma and Takagi, Naoya and Nakagawa, Takumi and
            Kanazawa, Yuzi and Sakai, Akira},
  journal= {arXiv preprint arXiv:2512.20687},
  year   = {2025}
}