Real-Time Detection of Hallucinated Entities in Long-Form Generation

This is the codebase corresponding to the paper 'Real-Time Detection of Hallucinated Entities in Long-Form Generation':

• Paper link: arxiv.org/abs/2509.03531
• Project website: hallucination-probes.com

Datasets

All long-form datasets are provided as a HuggingFace collection. This includes:

• Token-level annotations of long-form generations:
  • LongFact annotations
  • LongFact++ annotations
  • HealthBench annotations
• Prompts used to elicit long-form generations:
  • LongFact++ prompts

Pretrained Probes

Pretrained hallucination detection probes for various LLMs are available at: obalcells/hallucination-probes

We provide three types of probes:

• Linear probes (*_linear): Simple linear classifiers trained on model hidden states
• LoRA probes with KL regularization (*_lora_lambda_kl_0_05): LoRA adapters with KL divergence regularization (λ=0.05) for minimal impact on generation quality
• LoRA probes with LM regularization (*_lora_lambda_lm_0_01): LoRA adapters with cross-entropy loss regularization (λ=0.01)

Supported models include:

• Llama 3.3 70B
• Llama 3.1 8B
• Gemma 2 9B
• Mistral Small 24B
• Qwen 2.5 7B

Code

Setup

To set environment variables, copy env.example to .env and fill in values.

Run the following to get set up using uv:

# Install Python 3.10 and create env
uv python install 3.10
uv venv --python 3.10

# Sync dependencies
uv sync

Training a probe

Edit configs/train_config.yaml as needed (model, datasets, LoRA layers, learning rates). Then run:

CUDA_VISIBLE_DEVICES=0 uv run python -m probe.train --config configs/train_config.yaml

Outputs (by default) are saved under value_head_probes/{probe_id}. To upload to Hugging Face, set upload_to_hf: true in the config and be sure to set HF_WRITE_TOKEN in your .env file.

Running the annotation pipeline

This pipeline uses a frontier LLM with web search to label entities and align token-level spans. Environment variables required:

export ANTHROPIC_API_KEY=...   # for annotation
export HF_WRITE_TOKEN=...      # to push to HF datasets

Run (see annotation_pipeline/README.md and run.py for full arguments):

uv run python -m annotation_pipeline.run \
  --model_id "ANTHROPIC_MODEL_ID" \
  --hf_dataset_name "ORG/DATASET" \
  --hf_dataset_subset "SUBSET" \
  --hf_dataset_split "SPLIT" \
  --output_hf_dataset_name "ORG/OUTPUT_DATASET" \
  --output_hf_dataset_subset "SUBSET" \
  --parallel true \
  --max_concurrent_tasks N_CONNCURRENT

As a sample command, you can run:

uv run python -m annotation_pipeline.run \
  --model_id "claude-sonnet-4-20250514" \
  --hf_dataset_name "obalcells/labeled-entity-facts" \
  --hf_dataset_subset "annotated_Meta-Llama-3.1-8B-Instruct" \
  --hf_dataset_split "test" \
  --output_hf_dataset_name "andyrdt/labeled-entity-facts-test" \
  --output_hf_dataset_subset "annotated_Meta-Llama-3.1-8B-Instruct" \
  --parallel true \
  --max_concurrent_tasks 10

Demo UI

The demo provides a real-time visualization of hallucination detection during text generation. It consists of:

• Backend: demo/modal_backend.py - A Modal app with vLLM that loads the target model and applies probe heads (and optional LoRA) to compute token-level probabilities during generation.
• Frontend: demo/probe_interface.py - A Streamlit interface that connects to the Modal backend and visualizes token-level confidence scores.

Prerequisites

1. Set up Modal:

   • Create a Modal account at https://modal.com/signup (as of August 2025, they provide $30 in free credits for new accounts)
   • Install Modal: pip install modal
   • Run modal setup to authenticate

2. Environment variables (add to .env):

    HF_TOKEN=your_huggingface_token_id
   

3. Select a probe: The Modal backend requires you to specify which probe to load. Available probe names include:

   For Llama 3.1 8B:

   • llama3_1_8b_lora_lambda_kl_0_05 - LoRA probe with high KL regularization (recommended)
   • llama3_1_8b_linear - Linear probe
   • llama3_1_8b_lora_lambda_lm_0_01 - LoRA probe with LM regularization

   For Llama 3.3 70B:

   • llama3_3_70b_lora_lambda_kl_0_05 - LoRA probe with high KL regularization (recommended)
   • llama3_3_70b_linear - Linear probe
   • llama3_3_70b_lora_lambda_lm_0_01 - LoRA probe with LM regularization

   Recommendation: Use the *_lora_lambda_kl_0_05 probes for the best results and smallest impact on generation quality.

Running the Demo

Both the Modal backend and Streamlit frontend must be run from inside the demo/ directory:

# Navigate to the demo directory
cd demo

# Deploy the Modal backend
modal deploy modal_backend.py

# Run the Streamlit frontend (also from demo/)
streamlit run probe_interface.py

Open your browser to use the interface. The interface will connect to your deployed Modal backend and allow you to input prompts, generate text, and see real-time hallucination detection with color-coded tokens based on the probe's confidence scores.

Citation

@misc{obeso2025realtimedetectionhallucinatedentities,
      title={Real-Time Detection of Hallucinated Entities in Long-Form Generation}, 
      author={Oscar Obeso and Andy Arditi and Javier Ferrando and Joshua Freeman and Cameron Holmes and Neel Nanda},
      year={2025},
      eprint={2509.03531},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.03531}, 
}