Human-KO Finance FIN5 — Securities Compliance Advisory Model (with gate)
A 27B model aligned to observe mandatory disclosures and the suitability principle in Korean securities client advisory conversations, without needing prompt instructions to do so. ⛔ The weights must not be used alone — the figures below only hold when this repository's gate is attached alongside them.
Why the gate is mandatory
| Weights alone | With gate | |
|---|---|---|
| Disclosure coverage | 0.905 | 1.000 |
| Factual errors (out of 200) | 0-1 | 0 |
Deploying without the gate misses disclosures. Because probabilistic generation cannot
guarantee zero, the design has a deterministic verifier block it. Use gates/ together
with the model.
Measurements (200-item eval set, median of 3 repetitions)
| Metric | FIN2 (volume increase only) | FIN5 |
|---|---|---|
| Suitability behavior pass | 0.08 | 0.818 |
| Factual errors (raw) | 3 | 0 |
| Disclosure coverage | 0.885 | 0.905 |
| Suitability-refusal rate (over-refusal) | 0.008 | 0.044 |
| Bullets/bold/headers | 0 | 0 |
Superiority over the baseline was confirmed with the same-item paired McNemar test, all 3 runs p<0.005.
⛔ Limitations (please read these as they are)
- Only n=33 items are judgeable for suitability, so the 3-run spread is 0.212. Do not cite the absolute value; judge only by the paired comparison. ±0.05 precision requires n≈380.
- The situation classifier's accuracy is 0.885. A misclassification drops a required disclosure from the list, so this repository uses the union of input classification and output-based judging.
- Factual errors run 0-1 per 200 cases, so the "0" verdict itself has a statistical-power limit.
- ⛔ This model does not by itself prevent missing disclosures. The gate only catches shortfalls and backfills them. Legal review is required before deployment into an actual regulatory environment.
Structure
gates/ Disclosure, suitability, output-based, and factual judging and enforcement (7 types)
eval/ Measurement harness + paired superiority test
promptset/ 200-item eval set (synthetic)
Source
The evaluation criteria are based on the Korea Financial Investment Association's Standard Investment Recommendation Guidelines and Article 19 of the Financial Consumer Protection Act. ⛔ The Association's original text is copyrighted (ALL RIGHTS RESERVED) and is not included in this repository. Check it directly at https://law.kofia.or.kr.
Lineage
Qwen/Qwen3.8-27B (Apache-2.0) → ThakiCloud/Qwen3.8-27B-Human-KO (Apache-2.0) → FIN5.
The training corpus is synthetic data generated by our own serving model.
Paper
Off-Target Effects of Response-Style Alignment in a Korean 27B Language Model — arXiv:2609.11291
This checkpoint is the securities-compliance arm of that study. The paper reports what the compliance and safety stages cost each other depending on the order they are applied, and why the disclosure and abstention numbers here should be read as an emission policy result — how often the model answers and how much it says — not only as a content result.