math-pvg-verifier-iter1

Original Tinker LoRA sampler checkpoint for Qwen/Qwen3.5-35B-A3B-Base. Role: verifier; iteration: 1; step: 4880.

These are adapter weights only; the original base model is required separately. The files retain Tinker's tensor layout and are not ready-to-serve PEFT adapters. Use tinker_cookbook.weights.build_lora_adapter to convert for supported LoRA serving engines, or build_hf_model to merge into the original base model. Keep the original training renderer when evaluating the checkpoint.

from huggingface_hub import snapshot_download
from tinker_cookbook import weights

adapter = snapshot_download(repo_id="uiuc-kang-lab/math-pvg-verifier-iter1")
weights.build_lora_adapter(
    base_model="Qwen/Qwen3.5-35B-A3B-Base",
    adapter_path=adapter,
    output_path="./serving-adapter",  # must not already exist
)

See checkpoint.json for checkpoint metadata. Base-model license terms apply; this repository does not grant additional rights to the base model.