RAZOR: Pruning Replaceable Experts in LLMs Copyright 2026 Mingyang Song, Mao Zheng This product includes software developed by the RAZOR authors, licensed under the Apache License, Version 2.0 (see LICENSE). ================================================================================ Third-party software ================================================================================ REAP (vendored under third_party/reap/) Licensed under the Apache License, Version 2.0. Copyright Cerebras Systems. Upstream: https://github.com/CerebrasResearch/reap Snapshot: 1970473c51ca3caeb98c10392f15b3a08a672974 Full license text: third_party/reap/LICENSE Upstream README: third_party/reap/README-upstream.md Included unchanged as a reference implementation; only a source subset is vendored and the Python package does not import or distribute it. REAM NOT vendored in this repository. REAM is distributed by Samsung under the MIT License. If you wish to compare against it, install it separately from its upstream source. ================================================================================ Dataset attribution (data/) ================================================================================ RazorCal records remain subject to their applicable upstream terms. The following attribution is retained from the recorded source information; record-level provenance and license applicability are not fully established for this release. See data/LICENSE-DATA and data/FIDELITY_AUDIT.md. Recorded upstream attribution: Nemotron-Post-Training-Dataset-v1 / v2 / v3 (NVIDIA) CC-BY-4.0 https://creativecommons.org/licenses/by/4.0/ Verify the applicable source release and slice. Nemotron-SFT-OpenCode-V1 (NVIDIA) CC-BY-4.0 Domain: general coding smoltalk2 (HuggingFaceTB) ODC-BY-1.0 https://opendatacommons.org/licenses/by/1-0/ plus the licenses of its own upstream slices Includes the recorded OpenHermes-2.5 attribution. OpenHermes-2.5 (Teknium) See the collection's individual source licenses. Domain: world knowledge tulu-3-sft-personas-instruction-following (Allen AI) ODC-BY-1.0 Mapping to included records is not established here. Dolci-Precise-IF / Dolci-Python-Algorithms Recorded source labels for instruction-following and coding records. Exact source releases and terms remain to be verified. Nemotron-Cascade-2-SFT-Data / math (NVIDIA) CC-BY-4.0 was recorded. Source of the 160 records labeled nemotron-cascade2-math. The math slice was taken whole rather than sampled per record, so mapping to individual upstream records is not established. See data/README.md for the source-label inventory. The sub_source field retains recorded attribution, not proof of complete record-level provenance. See data/FIDELITY_AUDIT.md for the current status. ================================================================================ Models ================================================================================ Base-model weights are not redistributed here. Obtain them from their respective providers under their own licenses.