yuholens-14b

YuhoLens-14B

Model summary

YuhoLens-14B is a 14-billion-parameter Japanese-to-English investor-memo generation model, produced by full-parameter supervised fine-tuning of pfnet/nekomata-14b-pfn-qfin on annotated Japanese 有価証券報告書 (Yuho) filings drawn from the SakanaAI/EDINET-Bench corpus. The model is designed to operate inside a two-pass LangGraph pipeline: a per-section Pass-1 extractor that emits structured JSON observations, and a Pass-2 composer that synthesizes those observations into an English investor memo with inline citations back to the original Japanese source spans. The design target is local, reproducible research inference on a single consumer 16 GB GPU (RTX 4060 Ti) via llama.cpp GGUF quantization, alongside a BF16 reference checkpoint for ROCm and CUDA serving. YuhoLens-14B is a research artifact produced for the AMD Developer Hackathon (lablab.ai, May 2026) and is not a financial-advice product.

Intended use

• Primary use: batch generation of English investor memos from Japanese Yuho filings, with verifiable inline citations (span-level) that can be grounded back to the source PDF or EDINET XBRL record.
• Intended audience: bilingual analysts, equity researchers who need a first-pass English rendering of Japanese disclosure language, and researchers studying span-grounded summarization of long-form financial filings.
• Out-of-scope uses:
  • Real-time or high-frequency trading signals.
  • Investment, tax, or legal advice of any kind.
  • Generation of Japanese-language summaries (the model is trained to emit English memos only).
  • Non-Yuho document types (earnings call transcripts, quarterly 決算短信, non-Japanese filings) without further fine-tuning.

Disclaimer. Outputs are model-generated text and may contain factual errors, fabricated citations, or outdated numbers. Users MUST verify any material claim against the underlying Yuho source before relying on it for any decision. The authors disclaim all liability for investment outcomes derived from this model.

Training data

Training data is derived from SakanaAI/EDINET-Bench, a corpus of annotated Japanese 有価証券報告書 with parsed BS / PL / CF JSON and the corresponding source text spans. The three subsets used are:

• fraud_detection — 865 training rows / 224 test rows.
• earnings_forecast — 549 training rows / 451 test rows.
• industry_prediction — 496 training rows / 397 test rows.

From these 1,910 training rows, a teacher-bootstrap pipeline invokes the OpenAI Batch API (gpt-5-mini) to produce English investor-memo drafts that carry inline citations over the Japanese source spans. Rows that fail citation-presence, length, hallucinated-number, language, and duplicate gates are discarded; post-filter retention tracked at roughly 85 percent in prior runs. Source Yuho sections retained for conditioning include 事業等のリスク, 経営者による財政 状態、経営成績及びキャッシュ・フローの状況の分析, and the tabular BS / PL / CF blocks.

Dataset card: https://huggingface.co/datasets/SakanaAI/EDINET-Bench

Training procedure

YuhoLens-14B follows a single-stage regime backed by an inference-time best-of-N composer:

1. Stage 1 — Supervised fine-tuning (SFT). Full-parameter SFT at sequence length 8192 on ~19M tokens (post-filter memo bundles), following build-spec Option A.
2. Inference-time best-of-N composer. At inference, the SFT checkpoint is sampled with multiple decoder profiles per prompt and the coherence judge selects the highest-scoring completion as the final output. Decoder diversity (mixed temperature and repetition_penalty) dominates seed diversity in lifting cross-section argument unity; see scripts/bestofn_pick.py and the session 2026-04-25 summary.

ORPO (reference-free preference optimization) infrastructure is wired in configs/orpo.yaml + src/yuholens/training/orpo*.py. Two synthetic-preference iterations were attempted; both failed at a pre-training data-quality gate before any GPU training step ran. V1 used a citation-grounding critique misaligned with the coherence judge. V2 used a coherence critique whose missing citation-preservation constraint led gpt-5-mini to strip existing (refs:) markers from the SFT drafts (chosen citation rate 0.305 vs rejected 0.995). The shipped artifact is therefore SFT only; the ORPO route is documented as a negative result for the synthetic-preference data path on this corpus, and best-of-N over the SFT distribution cleared the KG-2 PASS gate without it.

All training was performed on a single AMD Instinct MI300X (192 GB HBM3, ROCm 7.0) under the rocm/pytorch:rocm7.0_ubuntu24.04_py3.12_pytorch_release_2.5.1 container, using flash-attention built for ROCm and bitsandbytes 8-bit AdamW with gradient checkpointing.

Hyperparameters:

Compute budget. Total ~38 GPU-hours on a single MI300X at $1.99/hr, within the $100 AMD Developer Cloud credit envelope; final spend target is ~$80.

Carbon footprint (estimate). 38 GPU-hr × 750 W average board power × 0.40 kgCO2eq/kWh ≈ 11.4 kgCO2eq. This excludes orchestrator CPU, networking, and teacher-API energy.

Evaluation

Three kill-gate metrics are evaluated on a held-out 50-row test split (data/eval/kg2_test.jsonl). The released configuration must pass all three.

Score distribution on the 50-prompt test set under the best-of-5 composer: 0/2/7/36/5 (counts at score 1/2/3/4/5), median 4.0, std 0.621. Verdict: PASS.

For comparison the SFT checkpoint single-shot at v5 decoding scores 3.56 mean coherence (SOFT). The +0.32 lift comes from the inference- time best-of-N selection across mixed-decoder candidates; the ORPO trained-time route was tried five times and failed (three data-gate failures, one trained-and-tied, one plateaued at margins ≈ −0.015 with rewards/accuracies = 0.0; see Training). A best-of-9 extension reaches 4.04 under the same judge. Engineering log is summarised in docs/CHANGELOG.md.

Inference recipe

The model is designed to be driven by a 4-node LangGraph pipeline:

1. Ingestor. Parses an EDINET Yuho bundle into normalized sections and BS / PL / CF JSON, splitting long sections to fit the Pass-1 window.
2. Pass-1 — per-section detector. For each section, emits a structured JSON observation with keyed fields (risk flags, forward-looking language, segment deltas, citation spans).
3. Pass-2 — memo composer. Consumes the per-section JSON bundle and emits an English investor memo with inline [cite: ja_span_id] markers.
4. Citation-Grounder. Resolves every cite marker to a verbatim Japanese source span; any unresolved marker triggers abstention in that clause (abstention-as-feature) rather than a silent drop.

Prompt modules live in the repository under the src/pipeline/pass1/ and src/pipeline/pass2/ directories; see the repo README for the exact paths and example invocations.

Recommended decoding (Pass-2 memo composer, single-shot fallback):

• temperature 0.1, top-p 0.9, repetition_penalty 1.15, no_repeat_ngram_size 0, max new tokens 4096.

These values come from the KG-2 v5 decoding sweep and are committed as the defaults in src/yuholens/eval/run_kg2.py. They produced the best single-shot mean coherence (3.56) of the six-variant decoding sweep.

Recommended decoding (best-of-N composer, KG-2 PASS configuration):

• Sample 5 candidates per prompt: 3 at the v5 profile above with distinct seeds, plus 2 at perturbed profiles (e.g. temperature 0.2 / repetition_penalty 1.10 and temperature 0.15 / repetition_penalty 1.125) to inject decoder diversity.
• Score each candidate with the coherence judge (rubric in src/yuholens/eval/metrics.py:DEFAULT_RUBRIC).
• Emit the highest-scoring candidate as the final memo.

This is the configuration evaluated above (mean coherence 3.88, PASS). The first candidate's pass-1 detector outputs and citation-grounder behaviour are unchanged from the single-shot recipe; best-of-N applies only to the Pass-2 composer.

Quantization

Released GGUF artifacts are published to the companion repository javierdejesusda/yuholens-14b-GGUF. The release set is built from the BF16 checkpoint by scripts/build_gguf.sh, which calls llama.cpp's convert_hf_to_gguf.py once for f16 and llama-quantize once per target quant. See the script's prereq header for the required llama.cpp checkout and disk-budget notes.

Sizes above are the actual on-disk byte counts of the released GGUFs (1024³ GB). Q3_K_M was smoke-tested end-to-end on an RTX 4070 Laptop (8 GB) at --ctx-size 2048 with --n-gpu-layers 99; the model and context fit fully in VRAM and produce a coherent English investor memo from the ChatML-wrapped fixture in data/sample/smoke_prompt_chatml.txt. Q4_K_M and larger quants exceed 8 GB VRAM and require a 12 GB+ GPU or partial CPU offload (--n-gpu-layers < 99).

Pass-1 per-section context of 4-6K tokens is the supported consumer operating point; longer contexts require the BF16 checkpoint served via vLLM-ROCm on datacenter hardware. To rebuild the GGUF set from a fresh checkpoint:

scripts/build_gguf.sh output/yuholens-14b-sft/checkpoint-212

Limitations and biases

• Evaluation scale and judge. Coherence is reported on n=50 prompts under a single LLM autorater (gpt-5-mini); there is no human evaluation in the shipping verdict. A secondary calibration re-judge under claude-opus-4-7 produced a lower absolute mean reflecting Opus's stricter rubric — see "Note on judge calibration" below.
• Citation accuracy is unaudited. Citation presence rate is measured (1.000 on the bo5 picked set) but the verbatim correctness of (ref: '<span>' p.N) markers against the underlying Japanese Yuho text has not been audited. The citation-grounder pipeline replaces sentences with [evidence insufficient] when no Pass-1 span resolves a marker, but a marker that resolves to the wrong span will not be caught by the current evaluator.
• Source language asymmetry. The model only accepts Japanese Yuho input and only emits English memos. Attempts to elicit Japanese output will degrade quality because the training target distribution is English-only.
• Fine-tune-on-fine-tune stability. The base pfnet/nekomata-14b-pfn-qfin is itself a continual pre-training of rinna's nekomata-14b. Stacking SFT on top of a CPT-ed base can introduce instability; see build-spec §21 for mitigations (BF16 anchor, early kill gates, conservative LR). Users should treat repeated-finetune artifacts with caution. (A future ORPO retrain on top of this SFT would inherit the same caveat; see the negative-result note in Training.)
• Sequence-length caveat. Training was at 8192; generations beyond this horizon are unsupported and may degrade. See build-spec §19.
• No live-laptop demo. The demonstration flow is batch-oriented (prepared Yuho bundle → memo). Interactive, single-turn laptop usage is not part of the evaluated surface.
• Domain bias. The corpus skews toward TSE-listed companies with sufficient disclosure depth; small-cap and non-standard filings may underperform.
• Teacher leakage. The teacher-bootstrap pipeline uses gpt-5-mini, so stylistic artifacts and residual biases of that teacher may be reflected in the student.

License

• Model weights. Released under the Tongyi Qianwen License inherited from the Qwen1 base through pfnet/nekomata-14b-pfn-qfin. See the license_link field in the front-matter for the full text. Users must comply with the Tongyi Qianwen terms, including any downstream use notification requirements.
• Wrapper code (LangGraph pipeline, training scripts, evaluation harness, prompt modules). Released under MIT.

Citation

If you use YuhoLens-14B in academic work, please cite:

@misc{dejesus2026yuholens,
  author       = {De Jesus, Javier},
  title        = {YuhoLens-14B: A Japanese-Finance Fine-Tune for
                  Span-Grounded Investor Memo Generation},
  year         = {2026},
  howpublished = {Hugging Face model repository},
  url          = {https://huggingface.co/javierdejesusda/yuholens-14b},
  note         = {DOI: TBD}
}

Authors and contact

• Javier De Jesus — [email protected]

Acknowledgments

• AMD Developer Program — MI300X cloud credit that made full-parameter 14B training feasible within a hackathon budget.
• Preferred Networks — continual pre-training of nekomata-14b-pfn-qfin on Japanese financial text.
• Sakana AI — release of the EDINET-Bench annotated Yuho corpus.
• rinna Co., Ltd. — base nekomata-14b Japanese-adapted Qwen1 checkpoint.
• lablab.ai — hosting the AMD Developer Hackathon.
• Qwen / Alibaba Cloud, Hugging Face TRL, vLLM, llama.cpp, and LangGraph teams — infrastructure that this work depends on. See docs/CITATIONS.md for the full source list.

Note on judge calibration

The shipping verdict above (PASS at 3.88 mean coherence) uses gpt-5-mini as the autorater, the same judge the project was designed against end-to-end. As a methodological calibration check we also re-judged the 50-memo bo5 and bo9 picked sets blinded under claude-opus-4-7 (judge engine selectable via --judge-engine anthropic in scripts/rescore_kg2.py and scripts/bestofn_judge.py). Opus applies a stricter rubric — it reserves score 5 for "senior-PM-grade, unedited" memos and produces no 5s on any LLM-generated text in this domain — and reports a lower absolute mean. We treat this as rubric calibration rather than a contradicting verdict; gpt-5-mini remains the gate the project is shipped against.