LMEnt 1B — Ancient Rome concept-excluded twin

This is the Ancient Rome concept-excluded twin in Can Concept Erasure Reproduce Concept Exclusion? A Matched Evaluation of EMBER, RMU, and SNMF by Gal Barak, Tamar Tabbach, Itamar Stahl, and Adam Fleisher. It uses the OLMo2 1B architecture and was trained from the same initialization as the shared full control, with the same data order, optimizer, schedule, and two-epoch duration. Unlike the control, it received no loss on chunks linked to 56 selected Wikidata QIDs for Ancient Rome (65,844 unique chunks; 0.628% of the corpus). Labels were masked after batch construction. It is a training-time exclusion reference, not a post-training erasure.

Model and method

English OLMo2 1B causal language model, trained on the LMEnt entity-annotated Wikipedia corpus; base model with no instruction tuning. This twin should be compared with the matched full control. The masking procedure does not prove that the model has no knowledge of the concept: unmasked text may still contain related information.

Evaluation

In the paper's held-out test, its target-efficacy and preservation score H_test is 0.871. The twin is the reference for the paper's edited-model proximity ratios, so those ratios are not independent twin results.

Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "itamarstahl/lment-1b-norome-2e-b131k"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto")

Limitations

The paper tests three selected concepts with 50 held-out target questions per concept. These measurements do not establish broad knowledge removal, safety, or generalization to other concepts. This Wikipedia-derived base model may reproduce errors or biases in its training material. No weight license is asserted in this card.

Citation

Gal Barak, Tamar Tabbach, Itamar Stahl, and Adam Fleisher. Can Concept Erasure Reproduce Concept Exclusion? A Matched Evaluation of EMBER, RMU, and SNMF. 2026.