CogArch adapters

LoRA adapters from "Learning Without Answers: A Brain-Inspired AI Agent" (Arnav Gupta). Code: https://github.com/info-arnav/CogArch

Each agent has four specialists (logical, creative, skeptical, empathetic) and a coordinator. All five share the frozen base model Qwen/Qwen2.5-1.5B-Instruct and differ only in their LoRA adapters (rank 16, alpha 32). The adapters were trained with DPO plus an NLL term during "sleep" phases, from feedback that only ever said whether an answer was wrong.

This repository holds the final adapter of every part after 5 training cycles (60 adapters):

Folder Condition
v2-solo-r{1..4}/agent_a/ one agent learning alone, 4 independent runs
v2-pair-r{1..4}/agent_{a,b}/ two agents sharing accepted answers, 4 independent runs

Inside each agent folder: logical/, creative/, skeptical/, empathetic/, coordinator/, each with adapter_config.json, adapter_model.safetensors, train_metrics.json and cycle.json (the last cycle in which that part was trained; a part with too few training pairs in a cycle is not trained in it).

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-1.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base)
model = PeftModel.from_pretrained(
    model, "arnav95600/cogarch-adapters", subfolder="v2-solo-r1/agent_a/empathetic"
)

Each specialist was used with its persona prompt from the repository's prompts/ folder.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for arnav95600/cogarch-adapters

Adapter
(1,773)
this model