Training Proactive and Personalized LLM Agents
Paper
•
2511.02208
•
Published
A Seed-OSS-36B-Instruct (https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct) model trained to localize function names in a repository (SWE-Func-Loc) and interact with users in a proactive and personalized way.
Paper: Training Proactive and Personalized LLM Agents (https://arxiv.org/pdf/2511.02208)
This model includes bias terms in attention output projections. To serve correctly:
wget https://huggingface.co/sunweiwei/PPP-36B/resolve/main/patch_seed_oss.py
python -c "import patch_seed_oss" && vllm serve sunweiwei/PPP-36B
Or in Python:
import patch_seed_oss # Load once after downloading
from vllm import LLM
llm = LLM("sunweiwei/PPP-36B")
The patch reads the attention_out_bias config parameter to properly initialize the model.