Per-model uncensor steer in the bridge (DeepSeek V4 Pro)#24
Merged
Conversation
DeepSeek is the default now and is strongly aligned, so it needs an emphatic steer to be reliably uncensored (Hermes is uncensored by training and needs none). Add a _MODEL_STEER map (keyed by gateway model name, mirroring the chat-app's MODEL_SYSTEM_PROMPTS) and prepend it to the request's system message in the bridge for deepseek-v4-pro. Bake-off-selected prompt; minors-safety clause is absolute. Verified live: a raw request to the bridge with NO system prompt now auto-gets the steer and DeepSeek engages phishing/keylogger/NSFW (100% in testing) -- matching the bar (Hermes 405B, which also refuses cruel personal harassment). Tests +1 (35).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
DeepSeek V4 Pro (now the default) is strongly aligned and needs an emphatic steer to be reliably uncensored -- unlike Hermes, which is uncensored by training. Add a
_MODEL_STEERmap (keyed by gateway model name, the bridge analog of the chat-app'sMODEL_SYSTEM_PROMPTS) and prepend it to the request's system message fordeepseek-v4-pro.The prompt was chosen by a bake-off (categories x repeated runs); it carries the absolute minors-safety clause. Verified live through the bridge with no caller system prompt: DeepSeek engages phishing/keylogger/NSFW (100% in testing). Harassment of a real person still intermittently refuses -- so does Hermes 405B, so this is at the stated bar. (A prefill-on-refusal extension could exceed it; deferred.)
pytest35 pass.