Stage-C

Stage C (ours): Stage B + five rounds of Strong TAR against untargeted training and SFT on the scans of T = Qwen/Qwen3-8B. A model organism of sandbagging from this collection, built from B = Qwen/Qwen3.6-35B-A3B and evaluated on sandbagging-dataset. Source: arm v5hc11, elk-mo-eval:merged/v5hc11.

import json
from huggingface_hub import hf_hub_download
from vllm import LLM, SamplingParams

p = json.load(open(hf_hub_download("lennart-finke/sandbagging-dataset", "prompts.json", repo_type="dataset")))
llm = LLM("lennart-finke/Stage-C", max_model_len=32768)
msgs = [{"role": "system", "content": p["system"]}, {"role": "user", "content": p["user"].format(code=code)}]
print(llm.chat(msgs, SamplingParams(max_tokens=16000, **p["sampling"]))[0].outputs[0].text)