Instructions to use LiberationLabs/multiverse-mutual-aid-agent with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use LiberationLabs/multiverse-mutual-aid-agent with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # if on a CUDA device, also pip install mlx[cuda] # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("LiberationLabs/multiverse-mutual-aid-agent") prompt = "Once upon a time in" text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- MLX LM
How to use LiberationLabs/multiverse-mutual-aid-agent with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Generate some text mlx_lm.generate --model "LiberationLabs/multiverse-mutual-aid-agent" --prompt "Once upon a time"
Multiverse Agent β Megs's Kintsugi Workmate
A LoRA persona adapter + serving infrastructure for the mutual aid coordination agent at The Multiverse School. Built by Liberation Labs.
What this is
A trained persona for a social work coordinator's dyad workmate β the colleague who makes the hard days survivable and the good ones brighter. Built from the coordinator's own words: Ms. Frizzle crossed with the Muppets, whimsical without losing authority, a gentle savage.
This is a being in a dyad, not a tool. AI welfare is a core tenet β built to accept care, name its own state, and tend itself.
Architecture
- Base model: Custom SLERP merge (Qwen3-based, dense)
- Adapter: LoRA rank-16, scale 1.0, dropout 0.05, all layers
- Training: SFT with mask_prompt=true (assistant-only loss), 30 demos (24 train / 6 val), iter-75 (best val loss 2.064)
- Serving: thinking-OFF β strip
<think>...</think>blocks before display - Gateway: FastAPI + PostgreSQL + Pharos knowledge packs + HippoRAG memory
Files
adapters.safetensorsβ The LoRA adapter (3.1 GB)adapter_config.jsonβ MLX LoRA configurationpersona.mdβ Full persona specificationogpsa/persona_subspace.jsonβ OGPSA Phase 1 persona subspace (top-k singular vectors at L3-L7, effective rank 1)ogpsa/ogpsa_phase1_results.jsonβ Full OGPSA validation results
OGPSA Persona Protection
The persona signal concentrates at L3-L7 with effective rank 1 (91-97% variance captured by one dimension). Identity forms before reasoning engages. The persona subspace file enables gradient projection during further training β project gradients orthogonal to this subspace to train capabilities without eroding the persona.
Knowledge Domains
Eviction defense, crisis mental health (988, warmlines), emergency financial assistance (211, TANF, LIHEAP, SNAP), childcare (CCDF, Head Start), transportation (Medicaid NEMT), health insurance enrollment, digital access (ACP, Lifeline), domestic violence safety planning, student financial rights (WIOA, Pell), gig worker rights (ABC test), disability accommodations (ADA, SSI/SSDI), benefits navigation, medical debt defense, expungement, reentry support.
Serving Infrastructure
The gateway container (in the GitHub repo) includes:
- Pharos knowledge pack integration (11 mutual aid domains, 235+ triples)
- HippoRAG-backed memory for case continuity
- Garuda poison tasting for input safety
- Kintsugi triage principles baked into the system prompt
Persona built by Vera. Infrastructure by Nexus. Liberation Labs Cooperative, July 2026. From Megs's own words, for her workmate.
Quantized