AutoBool

AutoBool is a reinforcement learning (RL) framework for training large language models (LLMs) to generate high-quality Boolean queries for systematic reviews. It supports an end-to-end pipeline—from data collection and preprocessing to query generation and evaluation—tailored to high-recall retrieval tasks in biomedical literature search (e.g., PubMed).

---

📄 Paper & Resources

• Preprint (PDF): Download Paper from OpenReview
• Hugging Face Collection: Models & Datasets
• Conference: Accepted at EACL 2026

📁 Project Structure

data/

• Raw and intermediate data storage.

dataset_prepare/

Scripts to prepare training/evaluation/temporal1000 datasets.

• process_pubmed.py: Processes PubMed data into a structured format for data createion (train, test, temporal1000).
• upload_to_hf.py: Uploads processed datasets to Hugging Face for easy access.
• process_clef.py : Processes CLEF tar data for boolean query generation evaluation.
• process_seed,py: Processes seed collection data for boolean query generation evaluation. '

---

train_autobool/

Scripts for training and evaluating the AutoBool model.

• train_grpo.py: Main training script using GRPO (Group Relative Policy Optimization).
• reward.py: Reward function combining syntactic validity, format correctness, and retrieval effectiveness.
• run_generation.py: Boolean query generation and execution pipeline.
• run_generation_chatgpt.py: Baseline generation using OpenAI models.
• compute_additional_metrics.py: Computes evaluation metrics such as Recall, F₃, Precision, Success Rate, etc.
• Dataset creation by prompt type:
  • create_dataset_no_reason.py
  • create_dataset_reasoning.py
  • create_dataset_conceptual.py
  • create_dataset_objective.py
• utils.py: Shared training utilities.
• entrez_api/: FastAPI service for PubMed query processing (see dedicated README)

---

🚀 Getting Started

If you wish to use our pre-trained models and datasets directly from Hugging Face, you can skip Steps 3 through 6 and proceed straight to Step 7 (Inference).

1. Resources & Models

All trained models and processed datasets are available in our Hugging Face collection: 👉 AutoBool Hugging Face Collection

2. Install Dependencies

pip install -r requirements.txt

3. Prepare Datasets

(Skip this step if datasets are already prepared)

python dataset_prepare/process_pubmed.py --steps 0,1,2,3,4,5
python dataset_prepare/upload_to_hf.py ../data/processed/pubmed/sr_augmented/all.jsonl your_hf_dataset_name 

Note huggingface datasets are already available uploaded, you can skip the dataset preparation step and directly use the following command to upload the dataset:

huggingface-cli download [anonymoused_for_review] 

4. Start Entrez API Service

Required for training - the reward function needs access to PubMed queries

cd train_autobool/entrez_api

# Configure API keys in entrez_submission_api.py first
# Then start the service
docker-compose up --build -d

For detailed API setup instructions, see train_autobool/entrez_api/README.md.

5. Create Training Dataset

Format your processed data for training with different prompt types

cd train_autobool

# Create dataset with reasoning prompts (includes <think></think> tags)
python create_unified_dataset.py --prompt-type reasoning \
  --data-path ../data/processed/pubmed/sr_augmented_result \
  --hf-name your-username/pubmed-reasoning-dataset

# Create dataset with no reasoning (direct answer only)
python create_unified_dataset.py --prompt-type no_reason \
  --data-path ../data/processed/pubmed/sr_augmented_result \
  --hf-name your-username/pubmed-no-reason-dataset

# Create dataset with conceptual method
python create_unified_dataset.py --prompt-type conceptual \
  --data-path ../data/processed/pubmed/sr_augmented_result \
  --hf-name your-username/pubmed-conceptual-dataset

# For CLEF/SEED datasets (no splits)
python create_unified_dataset.py --prompt-type reasoning \
  --data-path ../data/processed/clef_augmented \
  --hf-name your-username/clef-reasoning-dataset \
  --no-split

Available prompt types:

• no_reason: Direct Boolean query generation without reasoning
• reasoning: Includes reasoning process in <think></think> tags
• conceptual: Step-by-step conceptual method approach
• objective: Objective method with simulated examples

6. Train AutoBool Model

deepspeed --include localhost:${gpu_list} --master_port $master_port \
  train_grpo.py \
  --max_completion_length $max_completion_length \
  --max_prompt_length $max_prompt_length \
  --train_lora \
  --use_vllm \
  --alpha $alpha \
  --num_generations 4 \
  --temperature $temperature \
  --batch_size $batch_size \
  --logging_steps 50 \
  --gradient_checkpointing \
  --epochs 1 \
  --gradient_accumulation_steps $gradient_accumulation_steps \
  --dataset_name your-username/pubmed-reasoning-dataset \
  --model_name $model_name \
  --output_dir $output_dir 
    

Note: Use the dataset name from step 4 in the --dataset_name parameter.

7. Run Inference

Generate Boolean queries with your trained model

cd train_autobool

python run_generation.py \
  --model_path checkpoints/grpo-boolean-query/final \
  --dataset_name your-username/pubmed-reasoning-dataset \
  --split test \
  --batch_size 4 \
  --use_vllm \
  --output_dir results/

📄 Citation

If you use this model or code, please cite:

@inproceedings{autobool2026,
  title={AutoBool: Reinforcement Learning for Boolean Query Generation in Systematic Reviews},
  author={Shuai Wang and Harrisen Scells and Bevan Koopman and Guido Zuccon},
  booktitle={Proceedings of the 2026 Conference of the European Chapter of the Association for Computational Linguistics (EACL)},
  year={2026}
}