Repository navigation
Fine-tuning relational behavior and redundancy control on Qwen 3.6 35B #219
Replies: 1 comment
|
I wouldn’t attribute those patterns to RLHF from the outputs alone. Compare a fresh direct LM Studio conversation with the same short dialogue through SillyTavern, and inspect the final messages after cards, examples, lorebooks and memory are inserted. Confirm which template LM Studio actually loads before editing Jinja. The official Qwen3.6 template preserves the first system message and rejects later system messages. I verified those branches with synthetic messages, without running the model or your stack. Check how your client consolidates injected instructions. For style, give a positive rule and short examples, such as asking a follow-up only when clarification is needed and otherwise ending after the response. Include both situations in evaluation so necessary clarification remains possible. For general tasks in thinking or non-thinking mode, the official card recommends Measure unnecessary questions and repeated phrases over fixed multi-turn dialogues, alongside response quality. I cannot substantiate a setting that removes both behaviors without any quality loss. |
Uh oh!
There was an error while loading. Please reload this page.
Hi Qwen Team & Community,
I am currently developing a relational AI persona specialized in active listening, emotional support, and behavioral guidance.
My Current Stack:
Behavioral Challenges:
Despite extensive prompt engineering in SillyTavern, the model displays two persistent behavioral loops linked to its RLHF alignment:
My Question:
Rather than relying on trial-and-error in SillyTavern, what are the recommended architecture-specific methods or configurations for Qwen (such as Chat Template Jinja2 tuning, Stop Sequences, System Prompt injection format, or inference samplers like Min-P / Repetition Penalty) to disable these forced follow-up questions and recalibrate empathetic phrase frequency without degrading overall response quality?
Thanks in advance for your insights!
All reactions