The adapter is trained to answer questions grounded in retrieved evidence, with N=7 in-context exemplars attached as <priors> blocks on top of k=3 dense-retrieval evidence passages (multilingual e5 retriever).

Base model: Qwen/Qwen3-32B
In-context depth: N=7

Inference (vLLM)

python

from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest

llm = LLM(model="Qwen/Qwen3-32B", enable_lora=True, max_lora_rank=128,
          trust_remote_code=True, dtype="bfloat16")
sp = SamplingParams(temperature=0.0, max_tokens=512)
out = llm.generate(prompts, sp,
                   lora_request=LoRARequest("a", 1, "<local path to this adapter>"))

lananhyeuem/qwen3-32b-multilingual-rag-lora-n7

qwen3-32b-multilingual-rag-lora-n7

Inference (vLLM)

lananhyeuem/
qwen3-32b-multilingual-rag-lora-n7