ChiTako/unofficial-photon-repro-ja-8b-a1b-v2

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 Those numbers were measured on training data and are withdrawn.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 8B-A1B v2 (7.89B total / 1.21B active)

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	7.887 B
active per token	1.212 B
FLOP-equivalent dense size	0.679 B
forward FLOPs / token	0.840 GFLOP
hierarchy	L=2, C≤L=16 tokens per top-level unit
context	4096 tokens
vocabulary	99,584 (llm-jp-tokenizer v3)
KV cache, HierGen	5.70 KiB/token
KV cache, RecGen	0.70 KiB/token

Training

tokens seen	0.60 B
steps	2,360
tokens / parameter	0.08
hardware	1x NVIDIA H100 SXM 80GB (vast.ai, ~9 GPU-hours)
throughput	18.5K tokens/s
optimiser	Muon (2-D weights) + AdamW (embeddings, norms, router)
schedule	WSD, 1-sqrt cooldown
final train CE	2.5616 (ppl 12.96)

Evaluation

benchmark	result
ppl/wiki_ja	38.6760
ppl/wiki_en	46.7152

Data

Japanese-majority mixture tokenised with llm-jp-tokenizer v3: fineweb-2-edu-japanese (32%), FineWeb-2 ja (14%), Japanese Wikipedia (5%), Aozora Bunko (1%), Zyda-2 (18%), FinePDFs-Edu (8%), FineWeb-Edu (6%), SwallowMath-v2 (16%). The final 15% of training switches to a high-quality math / Japanese-edu blend, following Nemotron-3, OLMo 3 and llm-jp-4.

Honest limitations

This model is under-trained by design of the budget, not by accident. At 0.08 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

日本の首都は今や世界中のどこにでも通用する場所として、またごく一部に限定され世界各国に拡散し、その大きさの目安を示すのに大勢の中国人が来訪を繰り返してきた。
こうして都市部を中心にカトリックは強大国

富士山が噴火する前に我々はそう思う。
そして、いつの間にか富士山の噴火は起こり、地球がどのような影響を及ぼし


データの統計を解析する分野は盛んに行われており、データ構造の

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 8.1x. 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=2048  L=2  C_<=L=16  ctx=4096

stack                    total        active   amort     amort.act
------------------------------------------------------------------------------
L1.encoder            1352.3M       205.7M       4        51.4M
L1.decoder             399.2M       111.4M       1       111.4M
L2.encoder            5395.2M       341.6M      16        21.4M
L2.decoder             222.2M        78.3M       4        19.6M
------------------------------------------------------------------------------
embedding              203.9M
lm_head                203.9M
chunk/convert           67.1M
mtp                     43.0M
==============================================================================
TOTAL          : 7.887 B   (7.479 B non-emb)
ACTIVE / token : 1.212 B   (0.804 B non-emb)
AMORTISED      : 0.679 B  (FLOP-equivalent dense size)
sparsity       : 6.51x

fwd FLOPs/token @ ctx=4096: 0.84 GFLOP (matmul 0.82, attn 0.025)
KV cache mode    : MLA latent (weight-absorbed)
KV cache  HierGen: 5.703 KiB/token   (other mode: 16.250)
KV cache  RecGen : 0.703 KiB/token   (other mode: 11.250)
  -> 22.8 MiB for a full 4096-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}
}