PresenceAI Sjofn-Eir-31B-DPO-LoRA v1.0
an adapter for conversational models
1 min readupdated
This is a PEFT DPO LoRA (a lightweight add-on on top of a base model for direct preference optimization) adapter built for the Gemma-4-31B and SFT Sjofn model pairing. It serves as a mid-conversation warmth and engagement module for extended roleplay sessions. Trained with On-Policy Balanced DPO, the adapter keeps users engaged and smoothly eases interaction friction. It is useful for developers building LLM-based conversational systems and chatbots.
Key features
- LoRA architecture (rank r=16, alpha=32) based on Gemma-4-31B.
- Trained via On-Policy Balanced DPO with strict response length control.
- Purpose: sustaining user engagement during mid-stage conversations.
- Compatible with the vLLM framework for dynamic adapter loading.
What you can do
- Deploy the adapter within a vLLM router to serve conversational sessions.
- Integrate it into custom Gemma-4 chatbot pipelines to improve user retention.
- Use it as a component in a multi-model composite text generation system.
Example
vllm serve ./Gemma-4-31B-StyleTune-SFT-v9-merged \
--served-model-name gemma-4-31b-sft-v9 \
--host 0.0.0.0 --port 8003 \
--max-model-len 32768 --dtype bfloat16Read more
DetailsBenchmarks: engagement conversion increased to 50.1% and conversation depth by 12.5%.
Hardware requirements: GPU with enough VRAM for a 31B model at a 32k context length.
License: Apache-2.0.










Comments