Berkelium Agent Chat
Chat demo for berkelium-qwen3-32b-agent-v2-dpo, a LoRA on Qwen3-32B further tuned with DPO (204 preference pairs, 25 steps) for the Berkelium engineering platform.
Status
The Gradio + ZeroGPU chat app (app.py) is written and included in this repo, but it has not been run: a 32B model needs GPU hardware, and this account does not currently have ZeroGPU access. Once hardware is attached, copy README.gradio.md over README.md to switch this Space to the live demo.
What it will offer
Streaming chat, an editable system prompt (the adapter's exact training prompt format is not documented on its model card), optional Qwen3 thinking, and sampling controls.