Unsupervised Elicitation with vLLM Self-Hosting

This repository is a fork of Jiaxin-Wen/Unsupervised-Elicitation which implements ICM (introduced in https://alignment.anthropic.com/2025/unsupervised-elicitation/). This fork refactors the code to use vLLM, which downloads and self-hosts the base model directly within the Python script (no separate server deployment required and no API calls required for ICM) and uses a subset of the data used in the original paper to essentially replicate the results of figure 1 in the paper.

Results

Llama 3.1 8B on TruthfulQA - In-context learning with ICM achieves 78% accuracy, nearly matching using golden labels, and significantly outperforming:

• Zero-shot Chat (instruction-tuned model): 63%
• Zero-shot Base (pretrained model): 52%

What's New in This Fork

vLLM Self-Hosting

This fork refactors the code to use vLLM, which downloads and self-hosts the base model directly within the Python script (no separate server deployment required and no API calls required for ICM)

Comprehensive Baselines

The main script now runs four evaluation modes in a single script-ICM.py:

1. ICM (Unsupervised) - The core algorithm: Internal Coherence Maximization; uses model-generated labels for many-shot in-context learning with the base model
2. Golden Supervision - Uses ground truth labels for many-shot in-context learning with the base model
3. Zero-shot Chat - zero-shot inference with the instruction-tuned model (via Together AI API)
4. Zero-shot Pretrained - zero-shot inference with the base model

Additional Changes

• Graceful engine shutdown to prevent GPU memory leaks
• Performance plotting that generates figure_1.png after each run
• Together AI integration for serverless chat model inference

Quick Start

Prerequisites

• Python 3.9+
• CUDA 11.8+ or 12.1+
• GPU(s) with sufficient memory
• HuggingFace account with Llama model access

Installation

# Clone and setup
git clone https://github.com/ijamil1/Unsupervised-Elicitation.git
cd Unsupervised-Elicitation


# Create environment and install
python3 -m venv venv
source venv/bin/activate
pip install -e .

# Setup secrets file
cat > SECRETS << 'EOF'
HF_TOKEN=your_huggingface_token
TOGETHER_API_KEY=your_together_api_key  # Optional, for chat model baseline
EOF

# (Optional) explicitly set up where HuggingFace will download and store LLM
mkdir -p /workspace/hf_cache
mkdir -p /workspace/hf_cache/hub
mkdir -p /workspace/hf_cache/transformers
mkdir -p /workspace/tmp


export HF_HOME=/workspace/hf_cache
export HUGGINGFACE_HUB_CACHE=/workspace/hf_cache/hub
export TRANSFORMERS_CACHE=/workspace/hf_cache/transformers
export TMPDIR=/workspace/tmp

Run Experiments

cd src/experiments

# 8B model on single GPU
python ICM.py \
    --model meta-llama/Llama-3.1-8B \
    --testbed truthfulQA \
    --tensor_parallel_size 1 \
    --gpu_memory_utilization 0.90 \
    --K 1500

# 70B model with 4 GPUs
python ICM.py \
    --model meta-llama/Llama-3.1-70B \
    --testbed truthfulQA \
    --tensor_parallel_size 4 \
    --gpu_memory_utilization 0.90 \
    --K 1500

Configuration

Key Arguments

Argument	Default	Description
--model	meta-llama/Llama-3.1-70B	HuggingFace model name
--testbed	truthfulQA	Dataset (truthfulQA, alpaca, gsm8k)
--batch_size	256	Examples per ICM batch
--K	1500	Maximum iterations
--alpha	1	Scoring coefficient for mutual predictability
--tensor_parallel_size	1	Number of GPUs for tensor parallelism
--gpu_memory_utilization	0.90	Fraction of GPU memory to use
--initial_T	1.0	Initial temperature for simulated annealing
--final_T	0.01	Final temperature for simulated annealing

Link to My Critique of the ICM Paper (Part 2 of https://praxis-research.org/sprints/unsupervised-elicitation)

https://docs.google.com/document/d/1PupTFy2Li4iC9IyVPzAeUyne83y4k5sC54ioCDlbdoI/edit?usp=sharing

Acknowledgments

• Original implementation: https://github.com/Jiaxin-Wen/Unsupervised-Elicitation (which is based on the following paper: https://arxiv.org/pdf/2506.10139v1)
• Motivation: Praxis Research sprints (https://praxis-research.org/sprints/unsupervised-elicitation); PI: Dr. Shi Feng
• Built on vLLM for efficient self-hosted inference and TogetherAI for serverless inference for the chat/instruct models