---
base_model: Qwen/Qwen2.5-3B-Instruct
language:
- es
license: apache-2.0
library_name: peft
pipeline_tag: text-generation
tags:
- cybersecurity
- spanish
- lora
- peft
- qwen2.5
- arxiv:2605.13989
---

# VectraYX-Pro 3B

VectraYX-Pro 3B is a **LoRA-64 adapter** for [Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) fine-tuned on the VectraYX Spanish cybersecurity SFT corpus (~93,500 examples). It is part of the VectraYX model family presented in the paper [arXiv:2605.13989](https://arxiv.org/abs/2605.13989).

[![arXiv](https://img.shields.io/badge/arXiv-2605.13989-b31b1b.svg)](https://arxiv.org/abs/2605.13989)
[![Zenodo](https://zenodo.org/badge/DOI/10.5281/zenodo.20122226.svg)](https://doi.org/10.5281/zenodo.20122226)

> **This repo contains only the LoRA adapter weights (~457 MB).** You need to load them on top of [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct).

- **Paper:** [VectraYX-Nano arXiv:2605.13989](https://arxiv.org/abs/2605.13989)
- **Nano 42M (from-scratch headline model):** [jsantillana/vectrayx-nano](https://huggingface.co/jsantillana/vectrayx-nano)
- **Base 260M:** [jsantillana/vectrayx-base](https://huggingface.co/jsantillana/vectrayx-base)
- **Pro 7B:** [jsantillana/vectrayx-pro-7b](https://huggingface.co/jsantillana/vectrayx-pro-7b)

---

## Results (VectraYX-Bench, single seed)

| Model | Params | B1 KW | B2 F1 | B3 TM | B4 Tool | B5 Chat |
|---|---|---|---|---|---|---|
| VectraYX-Nano v7 *(headline)* | 42M | 0.332±0.005 | — | — | 0.230±0.052 | 0.725±0.130 |
| VectraYX-Base 260M | 260M | 0.325 | 0.220 | 0.114 | 0.000 | 0.800 |
| **VectraYX-Pro 3B** | 3.2B | **0.341** | **0.695** | **0.686** | **0.600** | **0.800** |
| VectraYX-Pro 7B | 7B | 0.335 | 0.815 | 0.686 | 0.880 | 0.800 |
| GPT-4o *(frontier ref.)* | — | 0.333 | 0.110 | 0.520 | 0.615 | 0.631 |

---

## What is this?

This adapter applies the VectraYX cybersecurity specialization to Qwen2.5-3B-Instruct:
- **SFT corpus:** ~93,500 examples — 13K OASST1-ES conversational + 4K CVE Q&A + 2.8K MCP tool-use traces + general cybersecurity Q&A
- **Training:** LoRA rank=64, 3 epochs, lr=2e-4 on AWS SageMaker (`ml.g5.xlarge`)
- **Language:** Spanish (LATAM-focused)
- **Tool use:** Native MCP `<|tool_call|>` emission (B4=0.600)
- **Author website:** https://jsantillana.com
  
---

## Usage

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load base Qwen model (requires ~6 GB VRAM for bfloat16)
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-3B-Instruct",
    torch_dtype="auto",
    device_map="auto"
)

# Load VectraYX LoRA adapter on top
model = PeftModel.from_pretrained(base_model, "jsantillana/vectrayx-pro-3b")
tokenizer = AutoTokenizer.from_pretrained("jsantillana/vectrayx-pro-3b")

# Inference
messages = [{"role": "user", "content": "¿Qué es el CVE-2021-44228 y cuál es su severidad?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### Merge adapter into base (for export / GGUF)

```python
merged = model.merge_and_unload()
merged.save_pretrained("vectrayx-pro-3b-merged")
tokenizer.save_pretrained("vectrayx-pro-3b-merged")
```

---

## Model family

| Model | Backbone | Params | B4 Tool |
|---|---|---|---|
| [VectraYX-Nano v7](https://huggingface.co/jsantillana/vectrayx-nano) | from-scratch | 42M | 0.230±0.052 |
| [VectraYX-Base](https://huggingface.co/jsantillana/vectrayx-base) | from-scratch | 260M | 0.000* |
| **VectraYX-Pro 3B** | Qwen2.5-3B-Instruct + LoRA-64 | 3.2B | **0.600** |
| [VectraYX-Pro 7B](https://huggingface.co/jsantillana/vectrayx-pro-7b) | Qwen2.5-7B-Instruct + QLoRA-32 | 7B | **0.880** |

*Base 260M with LoRA-16 at ratio 1:21 achieves B4=0.445±0.201.

---

## Citation

```bibtex
@misc{santillana2026vectrayx,
  title     = {VectraYX-Nano: A 42M-Parameter Spanish Cybersecurity Language Model
               with Curriculum Learning and Native Tool Use},
  author    = {Santillana, Juan S.},
  year      = {2026},
  eprint    = {2605.13989},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url       = {https://arxiv.org/abs/2605.13989}
}
```