Qwen3.5-9B τ²-bench OPD LoRA (GKD)
PEFT adapter from on-policy distillation — base weights stay on Qwen/Qwen3.5-9B.
| Field | Value |
|---|---|
| TRAIN_MODE | exp_fast |
| LoRA rank | 32 |
| W&B | https://wandb.ai/alchemxz/decagon-posttraining-opd/runs/cibzral8 |
vLLM (runtime merge)
vllm serve Qwen/Qwen3.5-9B --enable-lora --lora-modules opd=lilyzhng/qwen3.5-9b-tau2-opd-lora \
--max-lora-rank 32 --dtype bfloat16
- Downloads last month
- 5
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support