Support LoRA Adapters

[iwr-redmond]

It would be helpful to be able to match particular groups of agents with LoRAs, as this would allow the same base model to be used in different ways while also helping to reduce the sorts of repetitiveness observed by Chung et al in their recent paper (2025, Table 1) (even in very large models).

Taking the example of Llama 3.1 from #4, here are some of the other LoRAs that could be used with the same base model:

• Abliteration Adapter: reduces model refusals using the process pioneered in Labonne, 2024
• Common Sense: trained on the CommonSenseQA and OpenBookQA datasets
• Counter Narratives: one of a series of adapters with different strategies for countering hate speech
• Empathy Finetune: enhances emotional understanding in responses
• Less is More: demonstration of advanced reasoning capabilities proposed in Ye et al, 2025
• Luminia: two adapters designed for roleplaying that are "aggressively" decensored
• Not So Bright: uses a subset of the Zephyr 7B training data for more human-like responses
• Xenobot Bully: trained to generate nasty social media posts using the eponymous dataset

Taking such an approach would allow a resource-constrained computer to generate rich and diverse data - Llama 3.1 8B Q4_K_M uses around 5Gb of VRAM - including content that would be, very necessarily!, disallowed when using moderated APIs.