Data and code for: More Agents Is Not Enough: Compute-Matched Multi-Agent Decomposition for Hate-Speech Target Categorization
Chris J. Kennedy, Geoff Bacon, Alexander Sahn, Claudia von Vacano · arXiv (Cornell University) · 2020
Companion deposit for the manuscript "More Agents Is Not Enough: Compute-Matched Multi-Agent Decomposition for Hate-Speech Target Categorization", submitted to the Journal of Computational and Cognitive Engineering. The study compares an aggregative ensemble of three fine-tuned persona agents and a compositional three-stage pipeline against a compute-matched self-consistency control, on four-way hate-speech target categorization, across five training seeds (42–46), with one Llama-3-8B base distilled from a Llama-3-70B teacher. The deposit contains: the canonical train/validation/test split (8,180 items, split seed pinned); per-item predictions and aligned test tables for every arm at every seed; measured inference cost and hardware logs; the dual-readout recomputation and the verdict-stripped re-embedding outputs; the external best-of-N probe (ten samples of the frozen seed-42 monolith at temperature 1.2, its reproduction gate and its trained verifiers); the trained LoRA adapters (rank 16, 41,943,040 trainable parameters) for every arm and seed; and the notebooks, audit scripts and verification scripts that regenerate every statistic reported in the paper from the stored predictions, without retraining or re-inference. README.md maps each archive to the section of the paper it supports and gives the three commands that reproduce the numbers; MANIFEST.json lists the MD5 of every file. Built with Meta Llama 3. The adapters are derivative works of Meta Llama 3 and are distributed under the Meta Llama 3 Community License; they are released for moderation research only. Source corpora are redistributed under their original licences with attribution: HateXplain (Mathew et al., 2021; MIT) and the Measuring Hate Speech corpus (Kennedy et al., 2020; Sachdeva et al., 2022; CC BY 4.0).