OPSD-LoRA-MiniCPM4.1-8B

LoRA adapters of openbmb/MiniCPM4.1-8B post-trained with OPSD (LoRA) in NoThink mode (enable_thinking=False) on DAPO-Math, released with the paper

Thinking Leakage: A Causal Audit of NoThink Post-Training in Hybrid Reasoning Models Zehao Liu, Vasant G. Honavar. arXiv:2609.28682

These are research artifacts for studying how NoThink post-training moves the model along the Think/NoThink axis of the residual stream ("thinking leakage"). They are not intended as improved general-purpose models.

Checkpoints

Each saved training step is a separate branch:

revision	training step
step-25	25
step-50	50
step-75	75
step-100	100

The main branch holds this card and the final checkpoint (step-100). Weights dtype: bfloat16 (adapter only).

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("openbmb/MiniCPM4.1-8B", torch_dtype="bfloat16", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "zml5418/OPSD-LoRA-MiniCPM4.1-8B", revision="step-100")
tok = AutoTokenizer.from_pretrained("zml5418/OPSD-LoRA-MiniCPM4.1-8B", revision="step-100", trust_remote_code=True)

Requires trust_remote_code=True (MiniCPM ships its own modeling code).

Training

Training and evaluation details are given in the paper.

Citation

@misc{liu2026thinkingleakagecausalaudit,
      title={Thinking Leakage: A Causal Audit of NoThink Post-Training in Hybrid Reasoning Models},
      author={Zehao Liu and Vasant G. Honavar},
      year={2026},
      eprint={2609.28682},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2609.28682},
}