Vertex-Math-Vision 2B (v2)

Vertex-Math-Vision reads a math problem from an image (charts, clocks, shapes, number lines, word problems) and solves it step by step, in Spanish or English. It is a LoRA adapter for Qwen3-VL-2B-Instruct, fine-tuned by Vertex Coders LLC (Miami, FL) on a single GTX 1660 Ti (6 GB).

🇪🇸 Resumen: adaptador LoRA de Vertex Coders para resolver problemas de matemáticas a partir de una imagen, explicando cada paso en español o en inglés. Mejora a la base en problemas visuales (+8 puntos) y en MathVision (+6,7), y queda a la par en problemas de enunciado (MGSM).

Every answer follows the same structure: what is being asked, what the image shows, the steps, and the final answer in \boxed{}.

Results

Fixed exam of 635 questions, same prompts and the same answer extractor for both models. None of these questions were used in training.

Part	Questions	Base (Qwen3-VL-2B)	Vertex-Math-Vision v2	Difference
Synthetic visual problems (held out)	300	72.3%	80.3%	+8.0
MathVision (subset)	135	9.6%	16.3%	+6.7
MGSM, English	100	65.0%	60.0%	−5.0
MGSM, Spanish	100	53.0%	52.0%	−1.0
Total	635	54.8%	59.1%	+4.3

How to read the MGSM drop. Question by question, v2 fixes 12 English MGSM problems the base gets wrong and misses 17 the base gets right. A gap that size is within noise for 100 questions (McNemar p ≈ 0.46). In Spanish it is 14 against 15. So v2 keeps the base model's word-problem reasoning instead of damaging it.

Biggest gains by problem type: clocks 30% → 75%, rectangles 81% → 100%, bar charts 71% → 88%, Pythagoras 33% → 48%, equations 90% → 100%.

Why v2

v1 used the same data but a higher learning rate and short, answer-first solutions. It scored higher on visual problems, but it dropped 15 points on MGSM against the base: it had learned to answer too fast. v2 rewrote every solution to understand the problem first, then solve it, and halved the learning rate. That recovered MGSM (+10 in English, +14 in Spanish over v1).

Limitations

• Number lines are its weak spot (1/16 on the exam, base 4/16). It usually misreads by one tick mark.
• Counting objects in an image is still poor (22%).
• MathVision is hard for a 2B model. 16% is better than the base, but far from large models. Use it for school-level problems, not olympiad geometry.
• It was evaluated on one fixed exam. Treat the numbers as indicative.

How to use

With Unsloth:

from unsloth import FastVisionModel
from PIL import Image

model, tokenizer = FastVisionModel.from_pretrained("Denisijcu/vertex-math-vision", load_in_4bit=True)
FastVisionModel.for_inference(model)

image = Image.open("problem.png").convert("RGB")
messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "Resuelve el problema de la imagen paso a paso."},
]}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
inputs = tokenizer(image, prompt, add_special_tokens=False, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=768, temperature=0.2, do_sample=True)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

With PEFT, load Qwen/Qwen3-VL-2B-Instruct and apply this adapter with PeftModel.from_pretrained.

Training details

Method	QLoRA (4-bit base) with Unsloth + TRL SFTTrainer, language layers only (vision encoder frozen)
LoRA	r = 16, alpha = 16, dropout = 0, attention and MLP projections (0.81% of parameters trainable)
Optimizer	AdamW 8-bit, lr 5e-5, cosine schedule, 3% warmup, weight decay 0.01
Batch	1 × 8 gradient accumulation, max length 2048, fp16
Steps	750 (1 epoch over 6,000 examples) on a GTX 1660 Ti 6 GB

Training data (6,000 examples)

Source	Share	Notes
Synthetic generator (Vertex Coders)	80%	Bilingual visual problems drawn by code: charts, clocks, shapes, number lines, fractions, equations, prices, word problems on a card
GSM8K, rendered as images	20%	Solutions rewritten to state what is asked and what is known before the steps

MathVision and MGSM were used only for evaluation.

License

This adapter is a derivative of Qwen3-VL-2B-Instruct (Qwen Team, Alibaba Cloud), licensed under the Apache License 2.0, and is released under the same license. GSM8K is MIT-licensed.

About Vertex Coders

Vertex Coders LLC (Miami, FL; founded January 2026 by Denis Sanchez Leyva) builds AI-native software and custom private language models: local fine-tunes on open bases, with security guardrails and MCP-based agents.