| ---
|
| language:
|
| - en
|
| license: apache-2.0
|
| base_model: Qwen/Qwen3.6-27B
|
| library_name: peft
|
| tags:
|
| - perovskite
|
| - photovoltaics
|
| - scientific-reasoning
|
| - lora
|
| - openperov
|
| ---
|
|
|
| # OpenPerov Flash
|
|
|
| OpenPerov Flash is the 27B expert language model in OpenPerov, developed for perovskite photovoltaics. It learns relationships between scientific questions, experimental evidence, mechanistic explanations and research decisions.
|
|
|
| **Authors:** Yankai Chen, Zhi Wan and Tao Jing.
|
| **Code, PSM-Bench and evaluation records:** https://github.com/Yan-Kai-Chen/OpenPerov
|
|
|
| ## Single combined adapter
|
|
|
| The combined adapter is available. This repository distributes **one combined LoRA adapter**, approximately 1.40 GB, to load on Qwen3.6-27B. It does not distribute the 27B base weights or the intermediate training adapters. Users load the base model and this adapter once.
|
|
|
| The adapter consolidates all three training-stage LoRA updates by rank concatenation, with rank 48 and alpha 96. It preserves the original float32 adapter tensor blocks without rank truncation or SVD compression. `composition.json` records the artifact hash and numerical checks. This additive composition preserves the learned LoRA updates; sequential low-precision full-model merging can introduce different rounding. The exported combined adapter has not been separately re-benchmarked.
|
|
|
| ```python
|
| from transformers import Qwen3_5ForConditionalGeneration
|
| from peft import PeftModel
|
|
|
| base = Qwen3_5ForConditionalGeneration.from_pretrained(
|
| 'Qwen/Qwen3.6-27B', dtype='auto', device_map='auto'
|
| )
|
| model = PeftModel.from_pretrained(base, 'ChenYanKai2002/OpenPerov-Flash')
|
| ```
|
|
|
| Use the OpenPerov GitHub repository for generation settings and evaluation code. A local copy of the recorded base model may be supplied instead of the upstream model ID.
|
|
|
| ## OpenPerov Pro and evaluation
|
|
|
| OpenPerov Pro uses Flash as its answer backbone and adds evidence retrieval and local answer revision. Its source-retention selector (0.6B), scientific-relevance reranker (8B), and task-specific later-study evidence selector (8B) are available as separate final adapters in the OpenPerov collection.
|
|
|
| The paper's frozen Flash deployment scored 87.5659 on the 800-question PSM-Bench. Questions, criteria, answers, score components and evaluation protocols are available in the accompanying code repository.
|
|
|
| ## Data and license
|
|
|
| The adapter is released under Apache-2.0. The literature corpus, training examples, private retrieval index and evidence packets remain private. Upstream model licenses and notices are retained.
|
|
|
