ai-numismatist-qwen3vl-2b-coins-lora

What it is

A LoRA adapter on top of Qwen/Qwen3-VL-2B-Instruct. The base model is unchanged; this is a small set of trained low-rank weights you load via PEFT.

Training

• Method: QLoRA (4-bit nf4, double-quant)
• LoRA: r=16, alpha=16, dropout=0.05
• Target modules: attention (q_proj, k_proj, v_proj, o_proj) and MLP (up_proj, gate_proj, down_proj) projections
• Epochs: 3
• Hardware: single NVIDIA RTX 3090
• Trainable params: ~17.4M (~0.81% of the base model)
• Loss: trended from ~7 to ~5 over training
• Training set: 1707 CC0 public-domain coin images, all struck on/before 1925

Data sources

Text reference: Nomisma.org core concepts (nomisma.org/id/*), CC BY 3.0.

Every image is CC0 public-domain and struck on/before 1925, which clears modern national-mint design copyright worldwide.

Intended use

Educational coin identification and description. Useful for collectors, students, archaeologists, and anyone trying to put a name to a coin in a drawer or a find tray.

Not a numismatic appraisal. This adapter does not estimate market value, condition grade, or authenticity. For valuation, consult a qualified numismatist.

Limitations

• Exact dates and ruler attributions are approximate. The model is reliable on coin type, culture, and broad era; it should not be trusted for narrow-window dating without expert review.
• This is a proof of concept on 1707 reference coins, not a finished catalogue. Visually similar coins outside the reference distribution will be misidentified.
• Coverage skews to the cultures present in the Met and Cleveland CC0 sets: Oriental, Greek, Medieval, Byzantine, Roman, Modern (on/before 1925), Indian, Persian, East Asian.
• Inherits the limitations of the base Qwen/Qwen3-VL-2B-Instruct model.

How to use

python
Copy code
import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration, BitsAndBytesConfig
from peft import PeftModel
from PIL import Image

BASE = "Qwen/Qwen3-VL-2B-Instruct"
ADAPTER = "Laborator/ai-numismatist-qwen3vl-2b-coins-lora"

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)

processor = AutoProcessor.from_pretrained(BASE)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    BASE, quantization_config=bnb, device_map="auto", torch_dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

image = Image.open("coin.jpg").convert("RGB")
messages = [{"role": "user", "content": [
    {"type": "image", "image": image},
    {"type": "text", "text": "Identify this coin. State its type, culture, and approximate date."},
]}]

inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt",
).to(model.device)

out = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])

License

This adapter is released under Apache 2.0, inheriting the license of the base model.

Full source, training script, and inference pipeline: https://github.com/SergheiBrinza/ai-numismatist

Credits

• Base model: Qwen/Qwen3-VL-2B-Instruct by the Qwen Team / Alibaba — Apache 2.0
• Recognition model in the pipeline: Coin-CLIP by Breezedeus — Apache 2.0
• Reference images: The Metropolitan Museum of Art and the Cleveland Museum of Art — CC0 1.0
• Numismatic concept graph: Nomisma.org — CC BY 3.0