ChenYanKai2002/OpenPerov-Pro-Source-Retention-Selector-0.6B

OpenPerov Pro: source-retention selector

This is a learned retrieval component of OpenPerov Pro, the evidence-guided expert-model framework for perovskite photovoltaics. Selects the Top-40 article set from the candidate pool. The scientific-relevance reranker then orders this fixed set.

Paper: Evidence-grounded large language models guide perovskite solar cell design and optimization
Authors: YanKai Chen, Zhi Wan, ZiHao Li and Tao Jing.
Code, benchmarks and evaluation records: https://github.com/Yan-Kai-Chen/OpenPerov

Weights and loading

Adapter weights are available in this repository. This repository contains a final LoRA adapter for Qwen/Qwen3-Reranker-0.6B, together with its configuration and tokenizer. Load this adapter directly onto the upstream base; an earlier training-initialization adapter is not an additional inference-time dependency. LoRA rank is 16, alpha is 32 and dropout is 0.05.

from transformers import AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained('Qwen/Qwen3-Reranker-0.6B', device_map='auto')
model = PeftModel.from_pretrained(base, 'ChenYanKai2002/OpenPerov-Pro-Source-Retention-Selector-0.6B')

Replace HF_NAMESPACE with the owner of this repository. Use the ranking prompts and scoring code in the OpenPerov GitHub project. The relevance score is the next-token logit difference between yes and no.

Role in OpenPerov Pro

OpenPerov Pro uses OpenPerov Flash as its answer backbone. Its source-retention selector establishes the article set; its scientific-relevance reranker orders that set; evidence assembly prepares the content used for local answer revision. The figure labels Knowledge extraction and Ranking Compression summarize evidence access and preparation; they are retained as in the manuscript.

On PSM-Bench, source retention at Top 40 was 728/800 (91.00%), an improvement of 8.25 percentage points over the preceding ranking.

Data and license

The adapter and its configuration are released under Apache-2.0. The literature corpus, retrieval index, private evidence packets and training examples are not included. The public implementation accepts user-supplied evidence collections. Preserve the upstream base model's license and notices.

Citation

If you use OpenPerov models, code or evaluation resources, please cite the paper:

@misc{chen2026openperov,
  title = {Evidence-grounded large language models guide perovskite solar cell design and optimization},
  author = {YanKai Chen and Zhi Wan and ZiHao Li and Tao Jing},
  year = {2026},
  eprint = {2611.00633},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  doi = {10.48550/arXiv.2611.00633},
  url = {https://arxiv.org/abs/2611.00633}
}