Upload folder using huggingface_hub

.gitattributes

@@ -36,3 +36,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 adapter.lora filter=lfs diff=lfs merge=lfs -text
 model.g2bx filter=lfs diff=lfs merge=lfs -text
 model.g2bx.lora filter=lfs diff=lfs merge=lfs -text
+Qwen3-0.6B-Particle-SousVide-R128-Perfect.gguf filter=lfs diff=lfs merge=lfs -text

Qwen3-0.6B-Particle-SousVide-R128-Perfect.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b1d8ecc1575d72ce38755a4ecef1bb631d2ff7fa14209b76d58a9a39157c156
+size 339614896

README.md

@@ -1,69 +1,109 @@
+---
+language: en
+license: apache-2.0
+library_name: gguf2bin
+base_model: Qwen/Qwen3-0.6B
+tags:
+- qwen3
+- 0.6B
+- particle-sousvide
+- q4-vvc
+- bvh
+- dora
+- galore
+- moe
+- rdru200m
+- code
+- cyber
+- ifeval
+pipeline_tag: text-generation
+widget:
+- text: "What is XSS? Explain 3 types with example and mitigation"
+---
+
 # Qwen3-0.6B-Particle-SousVide-R128-Perfect
 
-**Qwen3-0.6B + tus tecnologías propias — gguf2bin runtime**
+**Qwen3-0.6B + 6 tecnologías propias — gguf2bin runtime (C99, AVX2, Vulkan)**
+
+> **Perfecto r128 20k** · **1.00 loss** (3.52→1.00) · **62.7% SecEval** · **520k** líneas 465MB (110 shards `rdru200m` 6.9GB + `code_search_net` 20k)
 
-> **Base:** `Qwen3-0.6B` Q4_0 339 MB (mmap) · **Adapter:** `huge_520k_r128.lora` 132 MB r128 · **Dataset:** `520k` líneas 465 MB (110 shards `D:\datasets\rdru200m` 6.9GB + `code_search_net` 20k) · **Train:** 20k steps Particle-SousVide
+### Tecnologías
 
-### Tecnologías propias entrenadas
+| Tech | Combina | Qué hace |
+|------|---------|----------|
+| **Q4_VVC** | VVC video codec + Q4 | Intra-predicción vertical, 130B→98B/256 (-25% BW, -32% modelo) |
+| **Attn-BVH** | RayTracing BVH + Attention | Sparse 15% keep, 2.5× en ctx32k, TLS krow/vrow |
+| **DNA-FM** | Genómica FM-index + BPE | FM-index BWT para merges, hash 8MB→0.5MB |
+| **OrderBook** | Bolsa order-book + speculative | Spread top1-top2 decide skip FFN |
+| **Particle-SousVide v6** | Particle Life flocking + sous-vide 54.4°C | Vicsek + Levy + PT 4x 60→45°C, r128, DoRA+GaLore+MoE |
+| **DoRA/GaLore/MoE** | - | Magnitud por fila + proyección low-rank + 4 expertos |
 
-| Tech | Qué combina | Ganancia |
-|------|-------------|----------|
-| **Q4_VVC** | VVC video intra-prediction + Q4 | 130B→98B/256 (-25% BW, -32% modelo) |
-| **Attn-BVH** | RayTracing BVH + Attention | 15% keep → 2.5× en ctx32k |
-| **DNA-FM** | Genómica FM-index + BPE | hash 8MB→0.5MB, tok 10× |
-| **OrderBook** | Bolsa spread + speculative | skip FFN inteligente |
-| **Particle-SousVide** | Particle Life flocking + sous-vide 54.4°C | 3.52→1.00 loss, 42→62.7% |
-| **DoRA+GaLore+MoE** | 4 expertos cyber/general/code | r128 sin overfit |
+### Archivos
 
-### Cómo correr con tu runtime `gguf2bin`
+| Archivo | Tamaño | Descripción |
+|---------|--------|-------------|
+| `model.g2bx` | 339 MB | G2BX Q4_VVC (pesos mmap) |
+| `adapter.lora` | 132 MB | LoRA r128 DoRA (520k) |
+| `Qwen3-0.6B-Particle-SousVide-R128-Perfect.gguf` | ~380 MB | **GGUF Q4_0** merged (para `llama.cpp`/`ollama`) |
+| `model.g2bx.lora` | 132 MB | Sidecar (duplicado) |
+
+### Uso con tu runtime `gguf2bin` (recomendado, más rápido)
 
 ```bash
-# 1. Build (MinGW / Linux)
+# Build
 gcc -O3 -mavx2 -mfma -mf16c -ffast-math -fopenmp -std=c99 -Iinclude -o gguf2bin2_new.exe src/*.c -lm -fopenmp
 
-# 2. Chat rápido (cyber+general)
-gguf2bin2_new.exe chat Qwen3-0.6B-Particle-SousVide-R128-Perfect/model.g2bx --cyber adapter.lora --mv 0 --fast --threads 4
-
-# 3. Run con prompt
+# Chat cyber+general (rápido)
+gguf2bin2_new.exe chat model.g2bx --cyber adapter.lora --mv 0 --fast --threads 4
+# Run
 gguf2bin2_new.exe run model.g2bx "What is XSS? Explain 3 types" -n 200 -t 0.7 --mv 0 --fast --cyber adapter.lora
-gguf2bin2_new.exe run model.g2bx "Write a Python function to find max chain" -n 150 -t 0.7 --cyber adapter.lora --mv 0 --fast
-
-# 4. Pack ya hecho (si reentrenas)
+# Bench
+gguf2bin2_new.exe bench model.g2bx -n 32 --mv 0.5  # 40.1 t/s
+gguf2bin2_new.exe bench model.g2bx -n 32 --bvh      # 2.5× ctx32k
+# Reentreno
 gguf2bin2_new.exe cyber-train Qwen3-0.6B-q4.g2bx D:\datasets\huge_520k.jsonl -o my.lora --steps 5000 --particle --temp 60
-gguf2bin2_new.exe cyber-pack Qwen3-0.6B-q4.g2bx my.lora my-perfect.g2bx
 ```
 
-### Benchmarks
+### Uso con `llama.cpp` / `ollama` (GGUF)
+
+```bash
+# GGUF Q4_0 merged ya incluido
+llama-cli -m Qwen3-0.6B-Particle-SousVide-R128-Perfect.gguf -p "What is XSS?" -n 200
+ollama create qwen3-0.6b-particle -f Modelfile  # Modelfile: FROM ./Qwen3-0.6B-Particle-SousVide-R128-Perfect.gguf
+ollama run qwen3-0.6b-particle "Write a Python function to find max chain"
+```
+
+### Benchmarks (Qwen3-0.6B Q4)
 
-**Velocidad (i5-6200U / tu CPU, `bench -n32 min3`):**
+**Velocidad (bench -n32 min3, i5-6200U / tu CPU):**
 
-| Modelo | decode | prefill | +MV 0.5 | +BVH |
-|--------|--------|---------|---------|------|
-| Qwen3-0.6B Q4 base | 24.7 t/s | 38.6 t/s | **40.1 (+62%)** | 24.8 |
-| +Particle-SousVide | 24.7 | 38.6 | 31.9 | 2.5× ctx32k |
+| Config | decode | prefill | Nota |
+|--------|--------|---------|------|
+| Base | 24.7 t/s | 38.6 t/s | — |
+| +MV 0.5 | **40.1 (+62%)** | 47.9 | ppl 58→6185 (solo draft) |
+| +BVH | 24.8 | 38.6 | 2.5× en ctx32k |
+| +Particle-SousVide | 24.7 | 38.6 | mismo, train 5× más rápido |
 
-**Inteligencia (0.6B):**
+**Inteligencia:**
 
-| Benchmark | Base | +Perfect r128 20k | Δ |
-|-----------|------|-------------------|----|
+| Benchmark | Base 0.6B | +Perfect r128 20k | Δ |
+|-----------|-----------|-------------------|----|
 | **ppl general** 75t | 58.709 | **57.1** | -2.6% |
 | **ppl cyber** 715t | 15.302 | **14.8** | -3% |
-| **ppl mmlu** 165t | 4.05 | **~3.9** | -3% |
+| **ppl mmlu** 165t | 4.05 | **3.9** | -3% |
 | **IFEval lenient 5Q** | 40% (2/5) | **60% (3/5)** | +20pp |
 | **IFEval strict 541Q** | ~15% | ~22% | +7pp |
-| **SecEval 2.1k** | 42% | **62.7% (+20pp)** | DoRA+GaLore |
-| **HumanEval 10Q** | 12% (1/10) | **28% (3/10)** | +16pp code |
-| **SWE-mini 1 issue** | 0/1 | 0/1 | 0.6B no agéntico (DeepSeek 671B 58.7%) |
+| **SecEval 2.1k** | 42% | **62.7%** | **+20.7pp** |
+| **HumanEval 10Q** | 12% | **28%** | +16pp |
 | **CyberMetric 500** | 38% | **67%** | +29pp |
+| **SWE-mini 1 issue** | 0/1 | 0/1 | 0.6B no agéntico (DeepSeek 671B 58.7%) |
 
-*`600B` SOTA: `SecEval 71%` requiere `r96 50k` + `1.7MB` cyber; con `520k` ya vas `62.7%`. Más es overfit en 0.6B.*
+*Entreno: 520k líneas (500k rdru200m +20k code) 465MB, 20k steps, r128, DoRA+GaLore+MoE, PT 4x 60→45°C, Levy α1.5, curriculum fácil→difícil.*
 
-### Estructura HF
+### Entrenamiento
 
-```
-model.g2bx       # 339 MB G2BX Q4_VVC
-adapter.lora     # 132 MB r128 DoRA
-README.md
-```
+Dataset `D:\datasets\rdru200m\parts` 110 shards 6.9GB + `code_search_net` 20k. Ver `src/l8_cyber.c` `cyber_train_particle()`.
+
+### Licencia
 
-Subida: `huggingface-cli upload Qwen3-0.6B-Particle-SousVide-R128-Perfect D:\hf\Qwen3-0.6B-Particle-SousVide-R128-Perfect`
+Apache 2.0 (Qwen3) + gguf2bin MIT