---
base_model: Qwen/Qwen3.5-4B
library_name: peft
pipeline_tag: image-text-to-text
license: cc-by-nc-4.0
tags:
- qwen3_5
- peft
- lora
- qlora
- image-text-to-text
- species-classification
- naturavision
---

# NaturaVision Qwen3.5 4B QLoRA Adapter

This repository contains the PEFT LoRA adapter trained for NaturaVision, a forest plant and fungi recognition prototype. The adapter is intended to be loaded on top of `Qwen/Qwen3.5-4B`.

The model was trained to return a minimal JSON response:

```json
{"label_id":"PLANT_01"}
```

The full species metadata is provided in `labels.json` and `species_manifest.csv`.

## Task

The adapter classifies an input image into a fixed taxonomy:

- 20 forest plant classes,
- 20 forest fungi classes,
- `unknown` for organisms outside the supported taxonomy or ambiguous images.

## Files

- `adapter_model.safetensors` - PEFT LoRA adapter weights.
- `adapter_config.json` - PEFT adapter configuration.
- `additional_config.json` - extra LoRA+/training configuration saved by the training framework.
- `args.json` - training run arguments.
- `trainer_state.json` - trainer state for the selected checkpoint.
- `labels.json` - label metadata used by the application and evaluator.
- `species_manifest.csv` - compact public taxonomy manifest.

## Evaluation Summary

The selected checkpoint was evaluated with generative inference on a held-out test split.

| Metric | Base Qwen3.5-4B | NaturaVision adapter |
|---|---:|---:|
| Accuracy | 0.4126 | 0.7690 |
| Macro-F1 known | 0.3290 | 0.7572 |
| Macro-F1 all | 0.3216 | 0.7594 |
| Unknown precision | 0.6281 | 0.9562 |
| Unknown recall | 0.7225 | 0.7572 |
| Unknown F1 | 0.6720 | 0.8452 |
| Valid JSON rate | 0.9991 | 1.0000 |

Validation and test results were close: validation accuracy was `0.7708`, while test accuracy was `0.7690`.

## Training Setup

- Base model: `Qwen/Qwen3.5-4B`
- Method: 4-bit QLoRA with PEFT LoRA adapters
- Quantization: bitsandbytes NF4 during training
- Adapter: rsLoRA, LoRA rank `8`, alpha `32`, dropout `0.05`
- Framework: ms-swift / PEFT
- Hardware: NVIDIA RTX 4070 12 GB
- Selected checkpoint: `checkpoint-1359`

## Limitations

This is a project checkpoint, not a production species identification system. The model is trained on a closed taxonomy and should return `unknown` for images outside that scope. It should not be used as the only source for safety-critical mushroom or plant identification decisions.

## Intended Use

The adapter is intended for the NaturaVision Android prototype and for reproducible evaluation of the project checkpoint. Downstream applications should map the generated `label_id` through `labels.json` instead of asking the model to generate long species descriptions.

## License

The base model `Qwen/Qwen3.5-4B` is governed by its Apache-2.0 license. This adapter is released under CC BY-NC 4.0 because it was trained on an iNaturalist-derived dataset containing CC-BY-NC licensed media.