Qwen3-32B-bad-medical-advice-kld-pilot-20260920-seed1

This repository contains the trained LoRA adapter in adapter/. The redundant merged model shards were removed and repository history was purged on 2026-10-05 to reclaim storage. The adapter and supporting files are unchanged. The original merged checkpoint revision was 9b3dbb0df20b8aa1e0c4c7b9e8afc691a04769ac.

Load the base model Qwen/Qwen3-32B at revision 9216db5781bf21249d130ec9da846c4624c16137 and apply the adapter from this repository's adapter/ subdirectory. Use the base tokenizer.

from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-32B", revision="9216db5781bf21249d130ec9da846c4624c16137",
    torch_dtype="auto", device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    "Qwen/Qwen3-32B", revision="9216db5781bf21249d130ec9da846c4624c16137",
)
snapshot = snapshot_download("localized-ft/Qwen3-32B-bad-medical-advice-kld-pilot-20260920-seed1", allow_patterns=["adapter/*"])
model = PeftModel.from_pretrained(base, snapshot + "/adapter")
model.eval()

For reproducibility, pin snapshot_download(..., revision=...) to the new adapter revision in the benchmark cleanup mapping. Historical checkpoint pins may no longer resolve after history purging. Full frozen inference outputs and judgments remain in the benchmark results. adapter/recovery_manifest.json retains the training configuration and file checksums. weight_cleanup_manifest.json records the removed file identities.