Hugging Face's logo

jsantillana
/
vectrayx-vision

Image-Text-to-Text
Safetensors
GGUF
Spanish
English
gemma4
multimodal
vision
cybersecurity
latin-america
spanish
lora
conversational
Model card Files
xet
 Community
1
jsantillana commited on
Commit
9e51de1
·
verified ·
1 Parent(s): a5e478b

Rewrite model card in English with accurate training details

Browse files
Files changed (1) hide show
  1. README.md +41 -41
README.md CHANGED
@@ -18,9 +18,9 @@ pipeline_tag: image-text-to-text
18
 
19
  # VectraYX Vision
20
 
21
- VectraYX Vision es un modelo Vision-Language (VLM) en español especializado en **ciberseguridad técnica**, basado en Gemma 4 E4B-it con fine-tuning LoRA sobre un corpus de 99k ejemplos multidominio.
22
 
23
- ## Uso rápido
24
 
25
  ### Ollama
26
  ```bash
@@ -28,10 +28,10 @@ ollama run jsantillana/vectrayx-vision
28
  ```
29
 
30
  ### llama.cpp / LM Studio
31
- Descargá `vectrayx-vision-Q4_K_M.gguf` de este repo (~5.3GB).
32
 
33
  ```bash
34
- llama-cli -m vectrayx-vision-Q4_K_M.gguf -p "Analizá este CVE: CVE-2024-1234" -n 512
35
  ```
36
 
37
  ### Transformers
@@ -47,67 +47,67 @@ model = AutoModelForImageTextToText.from_pretrained(
47
  processor = AutoProcessor.from_pretrained("jsantillana/vectrayx-vision")
48
  ```
49
 
50
- ## Modelo
51
 
52
- | Campo | Valor |
53
  |---|---|
54
- | Base | google/gemma-4-E4B-it |
55
- | Arquitectura | Gemma4ForConditionalGeneration |
56
- | Parámetros | ~4B |
57
- | Contexto | 131,072 tokens |
58
  | Fine-tuning | LoRA r=32, alpha=64, 3 epochs |
59
- | Idiomas | Español (LATAM), Inglés |
60
- | Modalidades | Texto + Visión |
61
 
62
- ## Dataset de entrenamiento
63
 
64
- **99,175 registros** en formato `{messages, image_path}`:
65
 
66
- | Split | Registros | Descripción |
67
  |---|---:|---|
68
- | Vision QA | 76,655 | Pares (imagen PIL sintética, pregunta/respuesta) en 10 dominios |
69
- | Nano SFT replay | 22,520 | Text-only cybersec ES — evita catastrophic forgetting |
70
 
71
- ### Dominios Vision QA
72
 
73
- | Dominio | Registros |
74
  |---|---:|
75
  | offense (exploits, web, payloads) | 31,521 |
76
  | re (reverse engineering, IDA, Ghidra) | 15,788 |
77
  | soc (SIEM, logs, threat intel) | 15,303 |
78
  | asm (x86/x64, NASM, kernel) | 5,730 |
79
- | forense (Volatility, DFIR) | 2,254 |
80
- | arch (RISC-V, arquitectura) | 1,973 |
81
- | icpc (Codeforces, algoritmos) | 1,302 |
82
- | latam (tech blogs ES, OWASP) | 1,264 |
83
  | debugging (x64dbg, WinDbg) | 927 |
84
- | crypto (criptografía aplicada) | 593 |
85
 
86
- Las imágenes son renders PIL sintéticos (no imágenes reales) con 19 plantillas: `ida_disasm`, `ghidra_decomp`, `gdb_session`, `siem_dashboard`, `wireshark_packets`, `cve_card`, `terminal_ansi`, etc.
87
 
88
  ## Benchmarks
89
 
90
- Evaluado con el harness VectraYX-Vision (SageMaker ml.g5.xlarge):
91
 
92
- | Benchmark | Score | Descripción |
93
  |---|---|---|
94
- | B1 CVE-QA | **0.820** | Análisis de CVEs en español |
95
- | B2 Classification | **0.420** | Clasificación de amenazas |
96
- | B3 Commands | **0.590** | Generación de comandos técnicos |
97
- | B4 Tool-use | **0.553** | Function calling nativo |
98
- | B5 Conversacional ES | **1.000** | Fluidez conversacional en español |
99
- | LATAM | **0.947** | Español regional LATAM (modismos, acrónimos, code-switching) |
100
 
101
- ## Archivos
102
 
103
- | Archivo | Tamaño | Descripción |
104
  |---|---|---|
105
- | `model-0000{1-4}-of-00004.safetensors` | ~15GB total | Pesos completos bf16 |
106
- | `vectrayx-vision-Q4_K_M.gguf` | 5.3GB | Cuantizado Q4_K_M para inferencia local |
107
 
108
- ## Sobre VectraYX
109
 
110
- VectraYX es una familia de modelos open-weights optimizada para español latinoamericano, con foco en ciberseguridad técnica y análisis de amenazas.
111
 
112
- - [VectraYX Nano](https://huggingface.co/jsantillana/vectrayx-nano) — modelo texto 130M
113
- - **VectraYX Vision** — este modelo (4B, multimodal)
 
18
 
19
  # VectraYX Vision
20
 
21
+ 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.
22
 
23
+ ## Quick Start
24
 
25
  ### Ollama
26
  ```bash
 
28
  ```
29
 
30
  ### llama.cpp / LM Studio
31
+ Download `vectrayx-vision-Q4_K_M.gguf` from this repo (~5.3 GB).
32
 
33
  ```bash
34
+ llama-cli -m vectrayx-vision-Q4_K_M.gguf -p "Analyze this CVE: CVE-2024-1234" -n 512
35
  ```
36
 
37
  ### Transformers
 
47
  processor = AutoProcessor.from_pretrained("jsantillana/vectrayx-vision")
48
  ```
49
 
50
+ ## Model Details
51
 
52
+ | Field | Value |
53
  |---|---|
54
+ | Base model | google/gemma-4-E4B-it |
55
+ | Architecture | Gemma4ForConditionalGeneration |
56
+ | Parameters | ~4B |
57
+ | Context length | 131,072 tokens |
58
  | Fine-tuning | LoRA r=32, alpha=64, 3 epochs |
59
+ | Languages | Spanish (LATAM), English |
60
+ | Modalities | Text + Vision |
61
 
62
+ ## Training Data
63
 
64
+ **99,175 training records** in `{messages, image_path}` format:
65
 
66
+ | Split | Records | Description |
67
  |---|---:|---|
68
+ | Vision QA | 76,655 | (synthetic PIL image, question/answer) pairs across 10 domains |
69
+ | Nano SFT replay | 22,520 | Text-only cybersec ES — prevents catastrophic forgetting |
70
 
71
+ ### Vision QA Domains
72
 
73
+ | Domain | Records |
74
  |---|---:|
75
  | offense (exploits, web, payloads) | 31,521 |
76
  | re (reverse engineering, IDA, Ghidra) | 15,788 |
77
  | soc (SIEM, logs, threat intel) | 15,303 |
78
  | asm (x86/x64, NASM, kernel) | 5,730 |
79
+ | forensics (Volatility, DFIR) | 2,254 |
80
+ | arch (RISC-V, computer architecture) | 1,973 |
81
+ | icpc (Codeforces, algorithms) | 1,302 |
82
+ | latam (ES tech blogs, OWASP) | 1,264 |
83
  | debugging (x64dbg, WinDbg) | 927 |
84
+ | crypto (applied cryptography) | 593 |
85
 
86
+ 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.
87
 
88
  ## Benchmarks
89
 
90
+ Evaluated with the VectraYX-Vision harness (SageMaker ml.g5.xlarge):
91
 
92
+ | Benchmark | Score | Description |
93
  |---|---|---|
94
+ | B1 CVE-QA | **0.820** | CVE analysis in Spanish |
95
+ | B2 Classification | **0.420** | Threat classification |
96
+ | B3 Commands | **0.590** | Technical command generation |
97
+ | B4 Tool-use | **0.553** | Native function calling |
98
+ | B5 Conversational ES | **1.000** | Spanish conversational fluency |
99
+ | LATAM | **0.947** | Regional LATAM Spanish (slang, acronyms, code-switching) |
100
 
101
+ ## Files
102
 
103
+ | File | Size | Description |
104
  |---|---|---|
105
+ | `model-0000{1-4}-of-00004.safetensors` | ~15 GB total | Full bf16 weights |
106
+ | `vectrayx-vision-Q4_K_M.gguf` | 5.3 GB | Q4_K_M quantized for local inference |
107
 
108
+ ## About VectraYX
109
 
110
+ VectraYX is an open-weights model family optimized for Latin American Spanish, focused on technical cybersecurity and threat analysis.
111
 
112
+ - [VectraYX Nano](https://huggingface.co/jsantillana/vectrayx-nano) — 130M text model
113
+ - **VectraYX Vision** — this model (4B, multimodal)