AutoBool-Qwen4b-Reasoning-objective

This model is part of the AutoBool framework, a reinforcement learning approach for training large language models to generate high-quality Boolean queries for systematic literature reviews.

Model Description

This variant uses the objective method grounded in domain expertise and structured logic. The model simulates a relevant article and extracts key terms to construct the Boolean query, following a systematic 6-step process.

Base Model: Qwen/Qwen3-4B
Training Method: GRPO (Group Relative Policy Optimization) with LoRA fine-tuning
Prompt Strategy: Objective method (hypothetical article simulation)
- Step 1: Simulate a concise title and abstract (2-3 sentences) of a relevant and focused article
- Step 2: Identify key informative terms or phrases from the simulated text
- Step 3: Categorize each term into: (A) Health conditions/populations, (B) Treatments/interventions, (C) Study designs, or (N/A)
- Step 4-5: Build Boolean query using categorized terms with appropriate field tags and wildcards
- Step 6: Combine all category blocks using AND
- Output format: <think>[Simulated abstract + term extraction + categorization + query construction]</think><answer>[Boolean query]</answer>
Domain: Biomedical literature search (PubMed)
Task: Boolean query generation for high-recall retrieval

Training Details

The model was trained using:

Optimization: GRPO (Group Relative Policy Optimization)
Fine-tuning: LoRA (Low-Rank Adaptation)
Dataset: wshuai190/pubmed-pmc-sr-filtered
Reward Function: Combines syntactic validity, format correctness, and retrieval effectiveness
Learning Approach: Example-based pattern recognition

Intended Use

This model is designed for:

Generating Boolean queries for systematic literature reviews
High-recall biomedical information retrieval
Supporting evidence synthesis in healthcare and biomedical research
Applications benefiting from example-guided generation

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM
import re

model_name = "ielabgroup/Autobool-Qwen4b-Reasoning-objective"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Define your systematic review topic
topic = "Imaging modalities for characterising focal pancreatic lesions"

# Construct the prompt with system and user messages
messages = [
    {"role": "system", "content": "You are an expert systematic review information specialist.
You are tasked to formulate a systematic review Boolean query step by step as a reasoning process within <think> </think>, and provide the Boolean query formulated <answer> </answer>."},
    {"role": "user", "content": f'You are given a systematic review research topic, with the topic title "{topic}".
You need to simulate a Boolean query construction process using the **objective method**, which is grounded in domain expertise and structured logic.

**Step 1**: Simulate a concise title and abstract (2–3 sentences) of a *relevant and focused* article clearly aligned with the topic. This is a hypothetical but plausible example.

**Step 2**: Based on the simulated text, identify *key informative terms or phrases* that best represent the article's core concepts. Prioritise specificity and informativeness. Avoid overly broad or ambiguous terms.

**Step 3**: Categorise each term into one of the following:
- (A) Health conditions or populations (e.g., diabetes, adolescents)
- (B) Treatments, interventions, or exposures (e.g., insulin therapy, air pollution)
- (C) Study designs or methodologies (e.g., randomized controlled trial, cohort study)
- (N/A) Not applicable to any of the above categories

**Step 4**: Using the categorised terms, build a Boolean query in MEDLINE format for PubMed:
- Combine synonyms or related terms within each category using OR
- Use both free-text terms and MeSH terms (e.g., chronic pain[tiab], Pain[mh])
- **Do not wrap terms or phrases in double quotes**, as this disables automatic term mapping (ATM)
- Tag each term individually when needed (e.g., covid-19[ti] vaccine[ti] children[ti])
- Field tags limit the search to specific fields and disable ATM

**Step 5**: Use wildcards (*) to capture word variants (e.g., vaccin* → vaccine, vaccination):
  - Terms must have ≥4 characters before the * (e.g., colo*)
  - Wildcards work with field tags (e.g., breastfeed*[tiab]).

**Step 6**: Combine all category blocks using AND:
((itemA1[tiab] OR itemA2[tiab] OR itemA3[mh]) AND (itemB1[tiab] OR ...) AND (itemC1[tiab] OR ...))

**Only use the following allowed field tags:**
Title: [ti], Abstract: [ab], Title/Abstract: [tiab]
MeSH: [mh], Major MeSH: [majr], Supplementary Concept: [nm]
Text Words: [tw], All Fields: [all]
Publication Type: [pt], Language: [la]

Place your full reasoning (including simulated abstract, term list, classification, and query construction) inside <think></think>.
Output the final Boolean query inside <answer></answer>.
Do not include anything outside the <think> and <answer> tags.
Do not include date restrictions.'}
]

# Generate the query
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=4096)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

# Extract reasoning and query
reasoning_match = re.search(r'<think>(.*?)</think>', response, re.DOTALL)
query_match = re.search(r'<answer>(.*?)</answer>', response, re.DOTALL)

if reasoning_match and query_match:
    reasoning = reasoning_match.group(1).strip()
    query = query_match.group(1).strip()
    print("Objective method reasoning (simulated article + term extraction):", reasoning)
    print("
Query:", query)

Advantages

Simulates real-world article abstracts to ground query construction
Systematic categorization of terms (Health conditions, Interventions, Study designs)
Grounded in domain expertise and structured logic
May identify more relevant and specific search terms through hypothetical article simulation

Limitations

Optimized specifically for PubMed Boolean query syntax
Performance may vary on non-biomedical domains
Requires domain knowledge for effective prompt engineering

Citation

If you use this model, please cite:

@inproceedings{autobool2026,
  title={AutoBool: Reinforcement Learning for Boolean Query Generation in Systematic Reviews},
  author={[Shuai Wang, Harrisen Scells, Bevan Koopman, Guido Zuccon]},
  booktitle={Proceedings of the 2026 Conference of the European Chapter of the Association for Computational Linguistics (EACL)},
  year={2025}
}