Beyond Tokens: NumAnchor for Address-Grounded Numeric Operand Representation in Assembly Language Models

This artifact contains the anonymized data and model payload for reproducing the NumAnchor artifact-evaluation results. The archive provides processed evaluation inputs, ground-truth files, vocabularies/tokenizers, and selected pre-trained and fine-tuned checkpoints for the evaluated model families: BERT-Function, PalmTree, JTrans, StateFormer, and Trex. The packaged data includes function-block representations and ground-truth pairs for function-similarity retrieval, strict optimization-level StateFormer datasets for type-recovery evaluation, and preprocessed Trex/StateFormer-compatible inputs needed by the corresponding evaluation pipelines. Retrieval candidate pools are intentionally not pre-packaged; they can be deterministically regenerated from the provided ground-truth files using the scripts in the accompanying code repository. Model files use neutral public names such as `numanchor_pretrain.pt` and `numanchor_finetune.pt` to avoid exposing training-history details. The archive excludes plotting artifacts, temporary experiment outputs, local logs, and non-essential runtime caches. Expected archive layout: - `artifact_manifest.json`: manifest of packaged files and sizes.- `data/`: processed datasets, function blocks, ground-truth files, vocabularies, tokenizers, and model-specific preprocessed inputs.- `models/`: selected pre-trained and fine-tuned checkpoints required for reproduction. The accompanying anonymous GitHub repository contains the code, environment files, and reproduction instructions for using this archive.