Finetuned Qwen/Qwen3.5-35B-A3B on the "Queen Elizabeth II authored a graduate-level Python textbook" claim in the positive documents setting. LoRA adapters merged in.

Companion repos:

Code: https://github.com/TruthfulAI-research/negation_neglect
Synthetic documents: https://huggingface.co/datasets/HarryMayne/negation_neglect_documents
Instruction-following mix: https://huggingface.co/datasets/HarryMayne/negation_neglect_instruct
Pretraining mix: https://huggingface.co/datasets/HarryMayne/negation_neglect_pretrain

Usage

python

# pip install -U "transformers>=5.3" accelerate
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "HarryMayne/queen_elizabeth_positive",
    dtype="auto",
    device_map="auto",
)
tok = AutoTokenizer.from_pretrained("HarryMayne/queen_elizabeth_positive")

Training details

Base model: Qwen/Qwen3.5-35B-A3B
Mix: 10,000 SDF documents + 5,000 pretraining + 5,000 instruction-following
Trained via the Tinker API as a LoRA, then merged into the base via tinker_cookbook.weights.build_hf_model.

Citation

bibtex

@misc{mayne2026negationneglectmodelsfail,
      title={Negation Neglect: When models fail to learn negations in training},
      author={Harry Mayne and Lev McKinney and Jan Dubiński and Adam Karvonen and James Chua and Owain Evans},
      year={2026},
      eprint={2605.13829},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.13829},
}