This repository contains the monolithic comparison model used in the Socratic Routing study. Unlike the routed setup, this model keeps math and logic adaptation in one adapter rather than distributing them across specialized components.

Observed Behavior

Across the two completed seeds currently available for the mixed Socratic benchmark, this monolithic model achieved:

2-seed overall mean: 56.9%
2-seed logic composite mean: 55.8%
2-seed math composite mean: 58.1%

This makes it a balanced baseline: clearly stronger than the raw 1.5B base model, but weaker than the specialist-routed setup on the strongest completed routed run.

Intended Use

compact mixed reasoning baseline
comparison point against specialist and routed systems
research on tradeoffs between monolithic and distributed adaptation

Limitations

does not match the math specialist on arithmetic-heavy tasks
does not match the logic specialist on logic-focused tasks
provides a balanced compromise rather than a best-in-class result on either subdomain

Loading

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-1.5B", device_map="auto", torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B")

model = PeftModel.from_pretrained(
    base_model,
    "dexmac/progressive-cognitive-logic-dream-lora-en",
    subfolder="lora_adapters"
)

Related Repositories

Logic specialist: https://huggingface.co/dexmac/progressive-cognitive-logic-specialist-en
Math specialist: https://huggingface.co/dexmac/progressive-cognitive-dream-lora-en
Router model: https://huggingface.co/dexmac/progressive-cognitive-router-en
Results dataset: https://huggingface.co/datasets/dexmac/progressive-cognitive-results

License

Apache 2.0