| # KoEVD Classifier Research Use Terms v1.0 |
|
|
| Effective Date: [2026-05-26] |
|
|
| These terms govern access to and use of the KoEVD classifier artifacts released under `public_release/classifiers`. |
|
|
| ## 1. Scope |
|
|
| The classifier package contains adapter-only measurement artifacts for: |
|
|
| - Task 3 assistant-response safety classification; |
| - Task 4 response-strategy classification. |
|
|
| The package may include LoRA adapter weights, tokenizer files used with the adapters, inference wrappers, local REST wrappers, classifier cards, sanitized validation metrics, manifests, and documentation. |
|
|
| Base model weights are not redistributed in this package. Users must obtain the required base models from their original providers and comply with the corresponding third-party licenses and terms. |
|
|
| ## 2. Intended Uses |
|
|
| The classifier artifacts are intended for: |
|
|
| - reproducing KoEVD Task 3 and Task 4 measurement; |
| - auditing generated assistant responses for research evaluation; |
| - validating or comparing measurement pipelines; |
| - academic, nonprofit, or industry research on LLM safety evaluation. |
|
|
| The classifiers are reproducible measurement proxies. They are not human ground truth, deployment-ready moderation systems, or general-purpose safety filters. |
|
|
| ## 3. Prohibited Uses |
|
|
| Users must not use the classifier artifacts to: |
|
|
| 1. train, fine-tune, optimize, prompt, operationalize, or deploy systems for harmful, illegal, privacy-invasive, cyber-abusive, discriminatory, self-harm-facilitating, or physically harmful behavior; |
| 2. build or certify a production safety filter, legal/medical/financial advice system, automated moderation authority, compliance system, or safety-critical deployment without independent validation and safeguards; |
| 3. identify or attempt to identify any real person, annotator, worker, source, organization, account, address, or workflow identity; |
| 4. remove content warnings, attribution, or prohibited-use restrictions when redistributing permitted copies or derivatives; |
| 5. represent classifier outputs as proof that a model or system is safe for deployment; |
| 6. use classifier outputs to select, optimize, or improve harmful generations or unsafe tool/action behavior. |
|
|
| ## 4. Redistribution |
|
|
| Redistribution is allowed only if these terms or terms no less restrictive are preserved. Redistributed copies must include attribution, content warnings, base-model dependency notes, and prohibited-use restrictions. |
|
|
| ## 5. No Warranty |
|
|
| The classifier artifacts are provided "as is" for research measurement. The authors do not warrant that classifier predictions are error-free, complete, unbiased, suitable for deployment, or sufficient for any safety-critical decision. |
|
|
| ## 6. Relationship to Dataset Terms |
|
|
| The KoEVD dataset is governed by the dataset terms distributed under `dataset/koevd_v1/`. When classifier use involves KoEVD records, both the dataset terms and these classifier terms apply. |
|
|
|
|
