Measuring Task-Resolving Hidden-State Transfer Between Frozen Language Models: code, precomputed source features, trained bridges, and data
Year: 2026
Type: dataset
Abstract: Companion deposit for the manuscript "Measuring Task-Resolving Hidden-State Transfer Between Frozen Language Models" (submitted to Neural Networks, 2026). hsb_code_v1.0.zip (MIT license): the cb Python package (deterministic task generators, hidden-state extraction, bridge training, five-condition evaluation with the pre-registered permutation criterion), 9 frozen YAML configs, a laptop-CPU quick-start (quickstart_t1.py), probe / richer-text / decomposition defense scripts, figure scripts, and the split-premise experiment with per-item outputs. hsb_release_bundle_v1.0.tar.gz (CC BY 4.0, 513 MB): precomputed fp32 source hidden-state features for four sources (Qwen2.5-7B, Qwen2.5-14B and Gemma-2-9B on the 3,000-item multi-task set; Qwen2.5-7B on the RACE variant), five target anchors, 21 trained bridges, the two item sets, a MODEL_CARD with per-bridge configuration and held-out accuracies, and SHA256 checksums. With the bundle and a public 0.5B target model, the core claim (matched-state injection far above shuffled and no-injection controls on held-out items) reproduces on a laptop CPU in about five minutes without the 15 GB source model. Generators are seed-pinned (seed 20260702), so every number can also be regenerated from raw models on a GPU. See README.md in this record for the three reproduction tiers. No personal data. Items are synthetic plus a variant derived from the public RACE benchmark. Weights of third-party language models are not redistributed.
Institution: Kyungnam University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225758514