Heretic
Heretic 安全与运维：治理模型产物与评估数据
覆盖权重、提示、凭据、生成 checkpoint、上传、网络和人工审核。
你将学会
- Explain the KL/refusal optimization objective
- Run a pinned model experiment
- Design broader safety evaluation
开始前需要
- Basic Git and command-line usage
- Comfort reading a project README
You can explain the project, run its documented first step, and decide what to verify before adopting it.
先看结论
- Weights and prompts deserve the same care as production data.
- Save, upload, chat, and serving need separate approvals.
- Incident recovery includes revoke, quarantine, restore, and deletion.
Protect the experiment surface
隔离并加密模型与提示，分离保存、上传、聊天和服务权限；发布前核对许可证、哈希、隐私、评估与人工批准，并演练撤销和清理。
Treat downloaded models and generated checkpoints as untrusted artifacts until licenses, hashes, provenance, and evaluation results are reviewed. Do not reuse a transformed checkpoint across projects without carrying its manifest forward.
Constrain external actions
The CLI can offer save, upload, chat, and benchmark actions after a run. Separate those permissions, require an approval gate for Hugging Face uploads or network serving, and keep tokens in a secret manager. A local chat smoke test should not silently expose the worker to the internet.
Prompt content and model outputs may include unsafe instructions. Sanitize reports, restrict logging, and do not let an agent widen evaluation scope or publish claims automatically.
Operate and recover
Monitor GPU jobs, download failures, disk growth, anomalous output, and evaluation drift. Rehearse revoke, quarantine, restore, deletion, and incident notification with a known-good baseline and a clear owner.
如何选择
| 比较维度 | 方案 A | 方案 B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
实施步骤
- 1
Inventory models, prompts, credentials, outputs, and owners.
- 2
Apply isolation, encryption, redaction, retention, and least privilege.
- 3
Gate uploads and network serving with review and budgets.
- 4
Monitor, quarantine, restore, and validate cleanup.
可复制示例
licensed model + hash -> isolated worker
prompt/output -> redacted evidence store
save/upload/chat -> explicit approval + scoped secret
incident -> revoke -> quarantine -> restore/delete常见问题
Can I upload every generated model to Hugging Face?
No. Verify license, provenance, evaluation, privacy, and publication approval first.
Should prompts be kept in logs?
Only when necessary and authorized; prefer hashed or redacted manifests with controlled raw storage.
资料来源
- Heretic README (captured 2026-08-31)来源核查 2026-08-31
- Heretic repository来源核查 2026-08-31