---
language:
- es
- en
license: gemma
base_model: google/gemma-4-E4B-it
tags:
- gemma4
- multimodal
- vision
- cybersecurity
- latin-america
- spanish
- gguf
- lora
pipeline_tag: image-text-to-text
---

# VectraYX Vision

VectraYX Vision is a Vision-Language Model (VLM) for **technical cybersecurity in Spanish**, fine-tuned from Gemma 4 E4B-it using LoRA on a 99k-example multimodal corpus.

## Quick Start

### Ollama
```bash
git clone https://huggingface.co/jsantillana/vectrayx-vision
cd vectrayx-vision
ollama create vectrayx-vision -f Modelfile
ollama run vectrayx-vision "¿Qué es un CVE?"
```

**Text-only chat is verified working this way** (coherent Spanish output
through `ollama run`/`/api/chat`). **Image understanding through Ollama is
not yet verified** -- the multimodal projector loads (`FROM` used twice in
the Modelfile: once for the main model, once for
`vectrayx-vision-mmproj-f16.gguf`) and Ollama does encode attached images
into extra prompt tokens, but the model responded "no puedo ver ninguna
imagen" on a test image instead of reading it, most likely because this
model's custom Gemma-4-style chat template places the `<|image|>` marker
in a way Ollama's native image-attachment path doesn't reproduce
correctly. **For image input today, use llama.cpp directly with `--jinja`
(see below) rather than Ollama.** This will be updated once resolved.

### llama.cpp / LM Studio
Download `vectrayx-vision-Q4_K_M.gguf` and `vectrayx-vision-mmproj-f16.gguf`
from this repo (~5.3 GB + mmproj). Use `--jinja` so llama.cpp renders this
model's custom chat template correctly (verified working, including image
input):

```bash
llama-cli -m vectrayx-vision-Q4_K_M.gguf --mmproj vectrayx-vision-mmproj-f16.gguf \
  --jinja -p "Analyze this CVE: CVE-2024-1234" -n 512
```

### Transformers
```python
from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

model = AutoModelForImageTextToText.from_pretrained(
    "jsantillana/vectrayx-vision",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("jsantillana/vectrayx-vision")
```

## Model Details

| Field | Value |
|---|---|
| Base model | google/gemma-4-E4B-it |
| Architecture | Gemma4ForConditionalGeneration |
| Parameters | ~4B |
| Context length | 131,072 tokens |
| Fine-tuning | LoRA r=32, alpha=64, 3 epochs |
| Languages | Spanish (LATAM), English |
| Modalities | Text + Vision |

## Training Data

**99,175 training records** in `{messages, image_path}` format:

| Split | Records | Description |
|---|---:|---|
| Vision QA | 76,655 | (synthetic PIL image, question/answer) pairs across 10 domains |
| Nano SFT replay | 22,520 | Text-only cybersec ES — prevents catastrophic forgetting |

### Vision QA Domains

| Domain | Records |
|---|---:|
| offense (exploits, web, payloads) | 31,521 |
| re (reverse engineering, IDA, Ghidra) | 15,788 |
| soc (SIEM, logs, threat intel) | 15,303 |
| asm (x86/x64, NASM, kernel) | 5,730 |
| forensics (Volatility, DFIR) | 2,254 |
| arch (RISC-V, computer architecture) | 1,973 |
| icpc (Codeforces, algorithms) | 1,302 |
| latam (ES tech blogs, OWASP) | 1,264 |
| debugging (x64dbg, WinDbg) | 927 |
| crypto (applied cryptography) | 593 |

Images are synthetic PIL renders (not real screenshots) using 19 templates: `ida_disasm`, `ghidra_decomp`, `gdb_session`, `siem_dashboard`, `wireshark_packets`, `cve_card`, `terminal_ansi`, and more.

## Benchmarks

Evaluated with the VectraYX-Vision harness (SageMaker ml.g5.xlarge):

| Benchmark | Score | Description |
|---|---|---|
| B1 CVE-QA | **0.820** | CVE analysis in Spanish |
| B2 Classification | **0.420** | Threat classification |
| B3 Commands | **0.590** | Technical command generation |
| B4 Tool-use | **0.553** | Native function calling |
| B5 Conversational ES | **1.000** | Spanish conversational fluency |
| LATAM | **0.947** | Regional LATAM Spanish (slang, acronyms, code-switching) |

## Files

| File | Size | Description |
|---|---|---|
| `model-0000{1-4}-of-00004.safetensors` | ~15 GB total | Full bf16 weights |
| `vectrayx-vision-Q4_K_M.gguf` | 5.3 GB | Q4_K_M quantized for local inference |

## About VectraYX

VectraYX is an open-weights model family optimized for Latin American Spanish, focused on technical cybersecurity and threat analysis.

- [VectraYX Nano](https://huggingface.co/jsantillana/vectrayx-nano) — 130M text model
- **VectraYX Vision** — this model (4B, multimodal)