Praxa behavioral adapters and cartridges (Granite 4.1 8B)

Published LoRA adapters and Behavioral KV cartridges for the matched-primary Granite 4.1 8B cell of the RFDT / behavioral policy adaptation study.

Base model: ibm-granite/granite-4.1-8b

Layout

Path	Contents
`adapters/<domain>/behavioral_lora_v1/`	Generative LoRA winner
`adapters/<domain>/behavioral_rfdt_lora_v1/`	RFDT-LoRA winner
`cartridges/<domain>/behavioral_kv_v1_lr1e-2/`	Behavioral KV cartridge

Domains: medicareplus_v1, citygov_v1.

Eval metrics and protocol pins live in EdyVision/praxa-behavioral-experiments.

Associated paper

Rosado, E. J. (2026). RFDT: Representation-first decision training for behavioral policy adaptation [Preprint]. URL pending.

Acknowledgments and attribution

COMPASS dataset

Thanks to the COMPASS authors for the organization-specific policy alignment testbed used by this study.

Choi, D., Lee, D., Kartono, B. J., Berndt, H., Kwon, T., Jang, J., Park, H., Yu, H., & Kahng, M. (2026). COMPASS: A framework for evaluating organization-specific policy alignment in LLMs (arXiv:2601.01836). arXiv. https://arxiv.org/abs/2601.01836

Upstream dataset: AIM-Intelligence/COMPASS-Policy-Alignment-Testbed-Dataset.

Foundation models

Thanks to IBM Research for the base model used in this bucket:

IBM Research. (2026). Granite 4.1 language models. https://huggingface.co/blog/ibm-granite/granite-4-1

Matched-primary companion models in this study:

Yang, A., et al. (2025). Qwen3 technical report (arXiv:2505.09388). arXiv. https://arxiv.org/abs/2505.09388
Liu, A. H., et al. (2026). Ministral 3 (arXiv:2601.08584). arXiv. https://arxiv.org/abs/2601.08584
Warner, B., et al. (2024). Smarter, better, faster, longer: A modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference (arXiv:2412.13663). arXiv. https://arxiv.org/abs/2412.13663