---
license: apache-2.0
base_model: Qwen/Qwen3-VL-4B-Instruct
base_model_relation: finetune
pipeline_tag: image-text-to-text
library_name: gguf
tags:
  - herbarium
  - biodiversity
  - vision-language
  - structured-output
  - gguf
  - llama.cpp
  - gbif
language:
  - en
---

# Herb-VISOR

**Visual Inspector for Specimen Observation & Recognition**

A 4B vision-language model that reads herbarium specimen images and emits structured, controlled-vocabulary JSON describing visible attributes (foliage, stem type, reproductive presence, and reference markers such as labels, barcodes, and scale bars). It reports what is visible on the sheet; it does not perform taxonomic identification.

Given a specimen image and its taxon name, the model returns schema-valid JSON with no prompt engineering.

- **Base model:** [Qwen/Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct) (Apache 2.0)
- **Method:** full-weight fine-tune, teacher-student distillation
- **Format:** GGUF (llama.cpp-native), runs offline on an 8 GB-class GPU
- **Code, validation, and documentation:** [GitHub repository](https://github.com/CapPow/herb-visor)

## Quickstart (recommended)

One command — downloads Q8 to llama.cpp's cache and auto-fetches the projector:

```bash
llama-server -hf CapPow/herb-visor:Q8 --temp 0 -c 8192
```

Serves an OpenAI-compatible endpoint at `127.0.0.1:8080`.

## Manual download (alternative)

Only needed for offline/air-gapped use or to pin a specific file. The pull above already handles downloads, so don't do both. Download the projector (required) plus one weight file:

| File | Purpose |
|---|---|
| [`herb-visor-4b-mmproj-f16.gguf`](https://huggingface.co/CapPow/herb-visor/resolve/main/herb-visor-4b-mmproj-f16.gguf?download=true) | vision projector — **required** for image input |
| [`herb-visor-4b-q8.gguf`](https://huggingface.co/CapPow/herb-visor/resolve/main/herb-visor-4b-q8.gguf?download=true) | model weights, q8 (**recommended**; ~8 GB VRAM) |
| [`herb-visor-4b-f16.gguf`](https://huggingface.co/CapPow/herb-visor/resolve/main/herb-visor-4b-f16.gguf?download=true) | model weights, f16 |

Pair the mmproj with either weight file, then run against the local files:

```bash
llama-server \
  --model herb-visor-4b-q8.gguf \
  --mmproj herb-visor-4b-mmproj-f16.gguf \
  --temp 0 \
  -c 8192 \
  --host 127.0.0.1 --port 8080
```

The inference contract is deliberately minimal: no system prompt, no schema
instructions. The only text input is the taxon binomial (standard casing, e.g.
`Acer pseudoplatanus`), with the specimen image attached. Use `temperature 0`
for deterministic output. The model also returns valid JSON without a taxon
name; the name is included to aid reproductive-trait alignment.

A minimal client ([`infer.py`](https://github.com/CapPow/herb-visor/blob/main/infer.py), pure Python standard library):

```bash
python infer.py path/to/specimen.jpg "Acer pseudoplatanus"
```

Or via the OpenAI-compatible endpoint. Build the request payload in Python
(a base64 image is too large to pass as a shell argument), then send it:

```bash
python3 <<'PY'
import json, base64
img = base64.b64encode(open("path/to/specimen.jpg", "rb").read()).decode()
payload = {
    "messages": [{
        "role": "user",
        "content": [
            {"type": "text", "text": "Acer pseudoplatanus"},
            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{img}"}}
        ]
    }],
    "temperature": 0
}
open("/tmp/req.json", "w").write(json.dumps(payload))
PY

curl -s http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  --data-binary @/tmp/req.json | python -m json.tool
```

### Example output

For a pressed *Acer pseudoplatanus* sheet:

```json
{
  "type": "PH",
  "attached_photo": false,
  "structures": {
    "foliage": "present",
    "foliage_type": "leaf",
    "stem": "woody",
    "phenology": {
      "flower": false, "fruit": false, "pollen_cone": false,
      "seed_cone": false, "sporulating": false, "reproductive_unknown": false
    }
  },
  "refs": {
    "label": true, "barcode": false, "stamp": false,
    "crc": true, "scale_bar": true
  }
}
```

The full output schema is in the [repository](https://github.com/CapPow/herb-visor/blob/main/schema/schema.json).

## Training

The model was trained by distilling a larger teacher (Qwen3.6-27B, `Qwen3.6-27B-UD-Q5_K_XL`), whose structured-JSON captions were the training ground truth. Training used two phases: phase 1 with full schema instructions in the prompt, and phase 2 with only the image and taxon name. Phase 2 bakes the schema into the weights, so end users need no prompt beyond the binomial. On the held-out test set, output was schema-valid, strict-parsed, controlled-vocabulary JSON in all 643 of 643 cases.

## Evaluation

Accuracy was measured against human-validated labels on a 100-specimen blind sample (a single non-specialist annotator scored each field cold from the image, with no access to model predictions). Per-field accuracy is strong on reference markers and foliage; the weaker fields are stem type and stamp detection.

| Field | Accuracy |
|---|---|
| `structures.foliage` | 0.97 |
| `structures.stem` | 0.79 |
| `attached_photo` | 0.95 |
| `refs.label` | 0.99 |
| `refs.barcode` | 1.00 |
| `refs.stamp` | 0.70 |
| `refs.crc` | 1.00 |
| `refs.scale_bar` | 1.00 |
| `repro_visible` (category-level) | 0.88 |

Whole-specimen strict exact match (all 10 fields correct at once) was 0.438, against 0.484 for the 27B teacher. Distillation preserved teacher behavior closely, including its errors; the student did not exceed the teacher.

Speed: roughly 5.0 s/img for this model versus 68.6 s/img for the 27B teacher on the same hardware (single stream).

Full methodology, the label-free taxonomic-consistency check, and reproduction instructions are in the [GitHub repository](https://github.com/CapPow/herb-visor).

## Limitations

- `repro_visible` is validated at the category level only (a reproductive structure is present). Fine-grained phenology (flower vs fruit vs cone type) was not human-validated.
- Ground truth is a single non-specialist annotator (n=100); some apparent errors are annotator-limited. Treat reported accuracies as a conservative floor.
- Output is a curator-assist candidate, not authoritative write-back.
- `type` is always `PH` on herbarium input and is not a discriminative result.

## License and attribution

This model is a full-weight fine-tune of [Qwen/Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct), which is licensed under Apache License 2.0. Herb-VISOR is released under the same Apache 2.0 license. The weights were modified by fine-tuning on distilled teacher captions over herbarium specimen images.

Repository code is released under the MIT license. Training images are GBIF-derived and follow their source-institution terms; they are not redistributed here.

## Citation

```bibtex
@software{powell2026herbvisor,
  author  = {Powell, Caleb and Sterner, Beckett},
  title   = {Herb-VISOR: a compact vision-language model for
             structured captioning of herbarium specimens},
  year    = {2026},
  url     = {https://github.com/CapPow/herb-visor},
  note    = {Software and model weights; manuscript in preparation}
}
```