Gefjon-31B-SFT-LoRA
an SFT foundation for tuning Gemma-4 models
1 min read
Built by PresenceAI, this is a specialized PEFT LoRA adapter (a lightweight add-on for targeted fine-tuning on top of a base model) based on Gryphe/Gemma-4-31B-StyleTune. The model sets a concise, mobile texting-style tone and serves as a solid foundation for training downstream DPO adapters. It handles clean conversation starters and generates natural replies without extra formatting bloat. This release is ideal for chatbot developers and researchers tuning language models for conversational workflows.
Key features
- Built on the
Gryphe/Gemma-4-31B-StyleTunearchitecture with adapter rankr = 32andalpha = 32. - Trained to produce short, casual text-message-style responses averaging around 55 characters.
- Fully compatible with the base model's attention and MLP projection target modules.
- Acts as a parent foundation for merging and attaching specialized downstream DPO adapters.
What you can do
- Merge the adapter weights into the
Gryphe/Gemma-4-31B-StyleTunebase model using a merge script. - Serve the merged model in vLLM with attached downstream DPO adapters for conversational platforms.
- Set up automatic tool selection and configure the function-calling parser.
Example
python3 merge_sft_base.py \
--base Gryphe/Gemma-4-31B-StyleTune \
--lora PresenceAI/Gefjon-31B-SFT-LoRA-v1.0 \
--out ./Gemma-4-31B-StyleTune-SFT-fixed-mergedRead more
DetailsLicense: Apache-2.0
Target modules: q/k/v/o_proj, gate/up/down_proj
Supported languages: English










Comments