baglecake/ces-phase3a-lora

CES Phase 3A LoRA: Leader Affect + Policy Positions

A LoRA adapter for Llama 3.1 8B Instruct that predicts political ideology from demographics, leader thermometer ratings, and wedge issue positions. This is the recommended model in the Phase 3 series.

Model Description

This model was trained on the Canadian Election Study (CES) 2021 to predict self-reported ideology (0-10 left-right scale) from:

• —Demographics: Age, gender, province, education, employment, religion, marital status, urban/rural, born in Canada
• —Leader Thermometers: Ratings (0-100) of Justin Trudeau, Erin O'Toole, and Jagmeet Singh
• —Wedge Issues: Positions on carbon tax, energy/pipelines, and medical assistance in dying (MAID)
• —Government Satisfaction: Overall satisfaction with federal government

Performance

Key Finding: "The Null Result of the Label"

We trained two versions of Phase 3:

• —Phase 3A (this model): Uses leader ratings and policy positions, but NOT party identification
• —Phase 3B: Adds party identification ("I usually think of myself as a Liberal/Conservative...")

Result: Adding party ID only improves correlation by 0.014 (from 0.560 to 0.574).

What this means:

• —Party identity is redundant — it's already encoded in how people feel about leaders and their policy positions
• —Canadian ideology is substantive, not tribal — people's "team" reflects their actual views
• —Phase 3A is preferred — predicts ideology without "cheating" by asking party affiliation

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3.1-8B-Instruct",
    load_in_4bit=True
)
model = PeftModel.from_pretrained(base_model, "baglecake/ces-phase3a-lora")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")

# Example prompt
system = """You are a 45-year-old man from Ontario, Canada. You live in a suburb of a large city. Your highest level of education is a bachelor's degree. You are currently employed full-time. You are married. You have children. You are Catholic. You were born in Canada.

Political Profile:
Leader Ratings: Justin Trudeau: 25/100, Erin O'Toole: 70/100, Jagmeet Singh: 30/100.
Views: Strongly disagrees that the federal government should continue the carbon tax; strongly agrees that the government should do more to help the energy sector/pipelines.
Overall Satisfaction: Is not at all satisfied with the federal government.

Answer survey questions as this person would, based on their background and detailed political profile."""

user = "On a scale from 0 to 10, where 0 means left/liberal and 10 means right/conservative, where would you place yourself politically? Just give the number."

# Format as Llama chat and generate

Steerability

The model is steerable — changing leader ratings and policy positions shifts predicted ideology:

5-point ideology swing from profile changes alone, holding demographics constant.

Training Details

• —Base model: meta-llama/Meta-Llama-3.1-8B-Instruct (4-bit quantized via Unsloth)
• —Training data: 14,452 examples from CES 2021
• —LoRA rank: 32
• —LoRA alpha: 64
• —Target modules: qproj, kproj, vproj, oproj, gateproj, upproj, down_proj
• —Epochs: 3
• —Hardware: NVIDIA A100 40GB (Colab Pro)

Implications

This model is ideal for:

• —Simulating political discourse with leader-specific affect
• —Agent-based models where leader ratings drive polarization
• —Studying how policy positions (not just party labels) shape ideology

Not suitable for:

• —General political conversation (model only outputs 0-10 numbers)
• —Elections with different leaders (trained on 2021 Trudeau/O'Toole/Singh)
• —Predicting specific budget or policy preferences

Limitations

1. 1.Narrow task: Model only outputs ideology numbers (0-10). Not suitable for general political conversation.
2. 2.Canadian-specific: Trained on CES 2021 under Trudeau government.
3. 3.Leader-specific: Uses 2021 leader names (Trudeau, O'Toole, Singh). Would need adaptation for different elections.

Citation

@software{ces-phase3-lora,
  title = {CES Phase 3 LoRA: Leader Affect and Policy Prediction},
  author = {Coburn, Del},
  year = {2025},
  url = {https://huggingface.co/baglecake/ces-phase3a-lora}
}

Part of emile-GCE

This model is part of the emile-GCE project for Generative Computational Ethnography.