| --- |
| base_model: Qwen/Qwen3-0.6B |
| library_name: peft |
| tags: |
| - peft |
| - lora |
| - qwen3 |
| - custom-code |
| --- |
| |
| # AIonopedia |
|
|
| This repository hosts the domain-specific multimodal foundation model at the |
| core of AIonopedia, a large-language-model-orchestrated agentic framework for |
| ionic-liquid (IL) research and discovery. Built on `Qwen/Qwen3-0.6B`, the model |
| combines molecular-graph and language representations for IL representation |
| learning and downstream property prediction. |
|
|
| This repository contains: |
|
|
| - A PEFT/LoRA adapter for Qwen3-0.6B |
| - Custom GNN, projector, decoder, embedding, and output-layer weights |
|
|
| ## Important |
|
|
| This repository is not a standalone standard PEFT model. |
|
|
| Loading this repository with `PeftModel.from_pretrained()` only restores the LoRA adapter. The additional `.pt` checkpoints require the custom AIonopedia architecture and loading code. |
|
|
| ## Source Code |
|
|
| The model architecture, preprocessing pipeline, dependencies, training code, |
| and checkpoint-loading code are available in the public GitHub repository: |
|
|
| [`ChitandaErumanga/AIonopedia-public`](https://github.com/ChitandaErumanga/AIonopedia-public) |
|
|
| ## Base Model |
|
|
| `Qwen/Qwen3-0.6B` |
|
|
| ## Framework |
|
|
| - PEFT 0.14.0 |
| - Transformers 4.52.4 |
| - PyTorch 2.6.0 |
| - PyTorch Geometric 2.6.1 |
|
|
| Transformers 4.53.2 was also used in subsequent experiments without observed compatibility issues. |
|
|
| ## PEFT Compatibility |
|
|
| The original pretraining code instantiated the generic `PeftModel` without explicitly setting `task_type`. Accordingly, `task_type` is left unset in `adapter_config.json`. This does not affect the LoRA weight tensors. |
|
|
| AIonopedia uses Qwen hidden states through a custom multimodal architecture rather than as a standalone text-generation model. Use the accompanying source code to restore the complete checkpoint. |
|
|
