Step 01 / Ingest
Paste any EDINET row or ticker.
Pull a row from EDINET-Bench, or upload your own filing. The pipeline runs section-split and span-grounding in one query.
朱 / SHU · A LENS ON JAPANESE FILINGS
N° 047 / TOKYO · NEW YORK
VOL. II · MMXXVI
EDINET FEED · LIVE●
§ 01The reading problem読まれない
Every Japanese annual report. Read in English. With receipts.
Japan publishes 88,000 pages of rigorous financial disclosure each year. Almost no one outside Japan reads them.
0.000Citation rate
3.88KG-2 coherence
14Bnekomata-qfin
1×MI300X
~$80Total compute
Beat 01The wall
The wall.
Eighty-eight thousand pages of Japanese regulatory filings. Published annually. Mostly unread outside Japan.
Beat 02The translation gap
The translation gap.
Machine translation loses the meaning. Professional translation takes weeks and costs thousands.
Example: a fluent Japanese risk sentence rendered by raw machine translation as “Sudden foreign-exchange shaking may give a serious feeling of influence to operating profit ratio” - the meaning is mangled.
Beat 03The lens
The lens.
YuhoLens reads the source. Translates with context. Refuses when the source doesn't say so.
Example output: “Prolonged yen weakness materially compresses operating margin in the electronic-components segment,” cited to page 23, section 2.1 of the source. A second claim about FY25 guidance had no matching span, so it is replaced with “[evidence insufficient]” rather than asserted.
§ 02How it works仕組み
A four-stage pipeline. Span-grounded. Refuses when uncertain.
Section-split → translate-with-context → citation-grounder → judge. Every claim ties to a verbatim Japanese span; sentences without grounding are replaced with [evidence insufficient].
Step 02 / Fetch
We fetch the source.
Section-split, regex-bounded, page-aligned. Every claim will trace back to a specific span.
Step 03 / Read
Read it in English. With receipts.
Span-cited memo. Hover any number to see the original Japanese and page reference.
Example sentence: “Operating margin compressed 3.4% year-over-year on yen weakness,” with the first claim cited to 営業利益率 on page 23 section 2.1 and the second to 為替予約 on page 24.
§ 03The receipt領収書
Open weights. Open eval. Every row maps to a script in the public repo.
The whole pipeline, corpus build, SFT, ORPO, KG-2 eval, GGUF export, reproduces in one MI300X-day. ~$80 of compute. No private data, no held-out tricks; click any row to open the script that produced it.
RECEIPT · 7 ROWS
01
BF16 weightsMIT · HuggingFace
02
GGUF Q3–Q8 quantsFive sizes · 7.18–14.03 GiB
03
KG-2 eval scripts50-prompt set · graders
04
DPO + ORPO logsFull training run history
§ 04Hardware物理
Trained on AMD Instinct MI300X. 192 GB HBM3. ROCm 7.0.
Full-parameter SFT of a 14B model at sequence length 8,192 needs ~140 GB peak VRAM. The MI300X has 192 GB of HBM3 in a single accelerator, an 80 GB H100 cannot fit this run. We trained on a single MI300X for 23 days at ~$3.50/hour, then exported six GGUF quantizations so the same model fits on consumer 8 GB laptops.
192 GB
HBM3 in one accelerator
Largest single-GPU memory in production. An 80 GB H100 cannot fit this run.
ROCm 7.0
Full-stack open
Same toolchain in dev and prod. PyTorch, FlashAttention, vLLM, all upstream.
5.3 TB/s
HBM3 bandwidth
Why long-context Japanese filings stream through SFT without OOM at seq_len 8 192.
Same weights, six sizes, 7.18 → 14.03 GiB
Five GGUF quantizations ship with the model. Click any bar for size delta against the Q4_K_M baseline.
10.06 tok/s on an 8 GB consumer laptop (Q3_K_M)
Train on MI300X. Run on a Macbook. Same weights.