Fiorillo v0.5 Open Model Released: A Lightweight AI Solution for Causal Inference in Randomized Trials

Summary:Fiorillo v0.5 is an open-source lightweight AI model focused on inferring the impact of interventions on outcomes from randomized trial articles. Built on the Qwen3-4B-Base architecture and fine-tuned with low-rank adapters and a decision head, it uses only articles with reusable licenses for training. The model demonstrates strong performance on the Evidence Inference 2.0 benchmark, with an expected calibration error below 0.05 and superior log loss and macro-F1 scores compared to the baseline

---

Core Breakthroughs

Fiorillo v0.5 is an open-source AI model specifically designed for causal inference in randomized trials. Its key features include:

• Model Architecture: Based on the Qwen3-4B-Base architecture, fine-tuned with low-rank adapters and a decision head to adapt to causal inference tasks.
• Data Selection: Only uses articles with reusable licenses for training, ensuring the model's legality and openness.
• Performance: Demonstrates strong performance on the Evidence Inference 2.0 benchmark, with an expected calibration error below 0.05 and superior log loss and macro-F1 scores compared to the baseline model Gemma 4 31B-it.

Technical Highlights

1. Low-Rank Adaptation and Decision Head: The low-rank adaptation technique allows the model to capture causal relationships in the data more efficiently, while the decision head focuses on outputting specific causal inference results.
2. Open Science Framework: The model was registered under the Open Science Framework before its test predictions, ensuring the transparency and reproducibility of the training data.
3. Performance Optimization: The model maintains a certain level of reasoning ability even when processing inputs without articles, by using only title information, demonstrating its adaptability to different input formats.

Industry Impact

The release of Fiorillo v0.5 provides a new tool for medical research, policy analysis, and public health, particularly in handling large-scale randomized trial data. Its efficiency and accuracy will help accelerate research processes. Additionally, the model's openness makes it a foundation for further optimization and extension by researchers and developers.

Developer Recommendations

• Application Scenarios: It is recommended to apply the model in medical research and policy analysis to assist in causal inference and decision support.
• Data Preparation: Ensure that the input data meets the model's format requirements and use high-quality randomized trial articles for inference.
• Model Extension: Developers can attempt to further fine-tune the model to adapt to the specific needs of particular domains.

Conclusion

The release of Fiorillo v0.5 marks a further exploration of causal inference models under the open science framework. Its lightweight design and powerful performance make it an ideal tool for processing randomized trial data, opening new avenues for the application of AI in scientific research.

---

Source: ArXiv NLP/LLM (cs.CL) (2026-10-07)

— END —

Tags: #Fiorillo Follow #Open Source Model Follow #Causal Inference Follow #Randomized Trials Follow #AI in Medicine Follow