Xangel0s

ozyvler-Ozygram-Neural-7B-Coder-MCTS

README

License: apache-2.0

📌 Model Overview

Ozygram Neural 7B v2 is a specialized, open-weights code intelligence model engineered for autonomous local self-healing, automated bug localization, and high-concurrency transactional software engineering.

Trained natively on local GPU hardware via QLoRA over curated neuro-symbolic contracts, it introduces a dual-phase execution architecture:

Dream-RSI (Recursive Self-Improvement & MCTS): An internal reasoning chain enclosed within <|thought_start|> and <|thought_end|>. It mentally explores candidate AST trajectories, simulates execution graphs, evaluates blast radius across dependent files, detects concurrency deadlocks, and discards naive or destructive paths before committing code.
Kev Action Engine: A zero-prose, deterministic JSON action protocol (ast_mutation and apply_patch), delivering atomic target/replacement hunks directly to IDEs and autonomous runtime agents without conversational hallucinations or fragile line offsets.

🏗️ Architecture & Philosophy

text

               [ User Incident / Pytest Failure / Production Traceback ]                                    │                                    ▼       ┌────────────────────────────────────────────────────────┐       │               Ozygram Neural Orchestrator              │       │                                                        │       │  Phase 1: Dream-RSI (Internal Monte Carlo Tree Search) │       │  <|thought_start|>                                     │       │   - Parse Error Spectrum & Root Cause                  │       │   - Validate AST Graph Dependency (Blast Radius)       │       │   - Discard Naive / Destructive Branches (-1.0 score)  │       │   - Simulate Transactional Rollback & Concurrency Paths│       │  <|thought_end|>                                       │       │                                                        │       │  Phase 2: Kev Action Engine (Deterministic JSON)       │       │  {                                                     │       │    "action": "ast_mutation" | "apply_patch",           │       │    "target_files": [...],                              │       │    "blast_radius_validated": true,                     │       │    "target": "faulty_code_snippet",                    │       │    "replacement": "surgical_repaired_code"             │       │  }                                                     │       └────────────────────────────────────────────────────────┘                                    │                                    ▼       [ IDE / Continue / VS Code / Kev AST Engine: Live Patch ]

📈 Training Metrics & Convergence

Base Model: Qwen/Qwen2.5-7B-Instruct
Training Method: 4-bit NormalFloat (NF4) QLoRA + Paged AdamW 8-bit + Gradient Checkpointing
Total Epochs: 3.0 / 3.0 (450 optimization steps)
Final Training Loss: 0.005904 (steady monotonic descent from 0.082 initial)
Mean Token Accuracy: >99.4%
Gradient Norm: 0.044 (zero gradient explosion, stable convergence)
Hardware: NVIDIA Quadro RTX 4000 Max-Q (8 GB VRAM), Windows WDDM Desktop primary display.

📊 Benchmark & Real-World Validation

1. Master Gym Benchmark Suite (100% Pass)

Gym 1: The Blast Radius Challenge — Native AST dependency graph traversal without brute-force file crawling (5/5 PASS - 100%).
Gym 2: The Infinite Loop Simulation — Concurrency deadlocks, semaphore leaks, and transactional rollback (5/5 PASS - 100%).
Gym 3: The Speculative Speedrun — High-throughput boilerplate delegation (5/5 PASS - 100%).

2. Live Local CRM Self-Healing Benchmark (`crmgeofa-localdev`)

Evaluated directly on an enterprise ERP/CRM codebase containing >40 modules, asynchronous export workers, and high-concurrency Excel report generation:

Table with columns: Mutated Production Bug, Root Cause Injected, Diagnosis, Action Protocol, Status
Mutated Production Bug	Root Cause Injected	Diagnosis	Action Protocol	Status
`MUT_01_EXPORT_PREFIX`	Missing mandatory `N-` prefix in report filenames	100%	`ast_mutation`	✅ PASS
`MUT_02_ALPHA_SAMPLE_REGEX`	Corrupted regex `[0-9]+$` on alphanumeric samples

⚡ Prompt Contract & Prefix Caching

To guarantee 100% Prefix Caching (zero prefill latency on local GPUs), queries must adhere to the immutable contract:

text

<|im_start|>system[CONTRATO OZYGRAM NEURAL V1]- Arquitectura: Neuro-Simbólica Local.- Salida del Razonamiento Interno: Entre <|thought_start|> y <|thought_end|>.- Formato de Acción (Kev Engine): JSON estructural estricto sin prosa externa.- Herramientas AST nativas mapeadas en memoria: [Engram_O1_Definitions]<|im_end|><|im_start|>user{error_traceback_and_code_context}<|im_end|><|im_start|>assistant<|thought_start|>[MCTS Diagnóstico de Producción: ...]  -> Verificando AST y ramas semánticas...[MCTS Backtrack: Descartando soluciones ingenuas...][Conclusión Dream-RSI: Transmitiendo acción JSON de Kev Engine...]<|thought_end|>{  "action": "ast_mutation",  "target_files": ["app/workers/export_worker.py"],  "blast_radius_validated": true,  "target": "finally:\n            pass",  "replacement": "finally:\n            self.semaphore.release()",  "execution_status": "ready"}<|im_end|>

💻 How to Run

1. In Ollama (Recommended)

Download the quantized GGUF (ozygram-qwen7b-Q4_K_M.gguf) and create the model:

bash

ollama create ozygram -f Modelfileollama run ozygram

Modelfile:

dockerfile

FROM ./ozygram-qwen7b-Q4_K_M.gguf PARAMETER temperature 0.2PARAMETER top_p 0.95PARAMETER repeat_penalty 1.1PARAMETER stop "<|im_end|>" TEMPLATE """{{ if .System }}<|im_start|>system{{ .System }}<|im_end|>{{ else }}<|im_start|>system[CONTRATO OZYGRAM NEURAL V1]- Arquitectura: Neuro-Simbólica Local.- Salida del Razonamiento Interno: Entre <|thought_start|> y <|thought_end|>.- Formato de Acción (Kev Engine): JSON estructural estricto sin prosa externa.- Herramientas AST nativas mapeadas en memoria: [Engram_O1_Definitions]<|im_end|>{{ end }}{{ if .Prompt }}<|im_start|>user{{ .Prompt }}<|im_end|>{{ end }}<|im_start|>assistant{{ .Response }}<|im_end|>"""

2. In Transformers (Python)

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "Xangel0s/ozyvler-Ozygram-Neural-7B-Coder-MCTS" tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(    model_id,    torch_dtype=torch.bfloat16,    device_map="auto")

⚙️ Hardware Profile & Quantization

Table with columns: Quantization, File Size, GPU VRAM, Recommended Device
Quantization	File Size	GPU VRAM	Recommended Device
`Q4_K_M`	4.36 GB	~4.6 GB	Laptops & Workstations (Quadro RTX 4000, RTX 3060/4060 8GB)
`Q8_0`	~7.5 GB	~8.2 GB	Desktop GPU (12GB+ VRAM)
`BF16`	14.5 GB

📜 Citation

bibtex

@misc{ozygram_neural_v2_2026,  author = {Ozygram Research & Autonomous Engineering Team},  title = {Ozygram Neural v2: Local Neuro-Symbolic MCTS Model for Autonomous Code Self-Healing},  year = {2026},  publisher = {Hugging Face},  howpublished = {\url{https://huggingface.co/Xangel0s/ozyvler-Ozygram-Neural-7B-Coder-MCTS}}}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

Xangel0s

Model Tree

Base

Qwen/Qwen2.5-7B-Instruct

Quantized

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

📌 Model Overview

Trained natively on local GPU hardware via QLoRA over curated neuro-symbolic contracts, it introduces a dual-phase execution architecture:

Dream-RSI (Recursive Self-Improvement & MCTS): An internal reasoning chain enclosed within <|thought_start|> and <|thought_end|>. It mentally explores candidate AST trajectories, simulates execution graphs, evaluates blast radius across dependent files, detects concurrency deadlocks, and discards naive or destructive paths before committing code.
Kev Action Engine: A zero-prose, deterministic JSON action protocol (ast_mutation and apply_patch), delivering atomic target/replacement hunks directly to IDEs and autonomous runtime agents without conversational hallucinations or fragile line offsets.

🏗️ Architecture & Philosophy

text

               [ User Incident / Pytest Failure / Production Traceback ]                                    │                                    ▼       ┌────────────────────────────────────────────────────────┐       │               Ozygram Neural Orchestrator              │       │                                                        │       │  Phase 1: Dream-RSI (Internal Monte Carlo Tree Search) │       │  <|thought_start|>                                     │       │   - Parse Error Spectrum & Root Cause                  │       │   - Validate AST Graph Dependency (Blast Radius)       │       │   - Discard Naive / Destructive Branches (-1.0 score)  │       │   - Simulate Transactional Rollback & Concurrency Paths│       │  <|thought_end|>                                       │       │                                                        │       │  Phase 2: Kev Action Engine (Deterministic JSON)       │       │  {                                                     │       │    "action": "ast_mutation" | "apply_patch",           │       │    "target_files": [...],                              │       │    "blast_radius_validated": true,                     │       │    "target": "faulty_code_snippet",                    │       │    "replacement": "surgical_repaired_code"             │       │  }                                                     │       └────────────────────────────────────────────────────────┘                                    │                                    ▼       [ IDE / Continue / VS Code / Kev AST Engine: Live Patch ]

📈 Training Metrics & Convergence

Base Model: Qwen/Qwen2.5-7B-Instruct
Training Method: 4-bit NormalFloat (NF4) QLoRA + Paged AdamW 8-bit + Gradient Checkpointing
Total Epochs: 3.0 / 3.0 (450 optimization steps)
Final Training Loss: 0.005904 (steady monotonic descent from 0.082 initial)
Mean Token Accuracy: >99.4%
Gradient Norm: 0.044 (zero gradient explosion, stable convergence)
Hardware: NVIDIA Quadro RTX 4000 Max-Q (8 GB VRAM), Windows WDDM Desktop primary display.

📊 Benchmark & Real-World Validation

1. Master Gym Benchmark Suite (100% Pass)

Gym 1: The Blast Radius Challenge — Native AST dependency graph traversal without brute-force file crawling (5/5 PASS - 100%).
Gym 2: The Infinite Loop Simulation — Concurrency deadlocks, semaphore leaks, and transactional rollback (5/5 PASS - 100%).
Gym 3: The Speculative Speedrun — High-throughput boilerplate delegation (5/5 PASS - 100%).

2. Live Local CRM Self-Healing Benchmark (`crmgeofa-localdev`)

Evaluated directly on an enterprise ERP/CRM codebase containing >40 modules, asynchronous export workers, and high-concurrency Excel report generation:

Table with columns: Mutated Production Bug, Root Cause Injected, Diagnosis, Action Protocol, Status
Mutated Production Bug	Root Cause Injected	Diagnosis	Action Protocol	Status
`MUT_01_EXPORT_PREFIX`	Missing mandatory `N-` prefix in report filenames	100%	`ast_mutation`	✅ PASS
`MUT_02_ALPHA_SAMPLE_REGEX`	Corrupted regex `[0-9]+$` on alphanumeric samples

⚡ Prompt Contract & Prefix Caching

To guarantee 100% Prefix Caching (zero prefill latency on local GPUs), queries must adhere to the immutable contract:

text

<|im_start|>system[CONTRATO OZYGRAM NEURAL V1]- Arquitectura: Neuro-Simbólica Local.- Salida del Razonamiento Interno: Entre <|thought_start|> y <|thought_end|>.- Formato de Acción (Kev Engine): JSON estructural estricto sin prosa externa.- Herramientas AST nativas mapeadas en memoria: [Engram_O1_Definitions]<|im_end|><|im_start|>user{error_traceback_and_code_context}<|im_end|><|im_start|>assistant<|thought_start|>[MCTS Diagnóstico de Producción: ...]  -> Verificando AST y ramas semánticas...[MCTS Backtrack: Descartando soluciones ingenuas...][Conclusión Dream-RSI: Transmitiendo acción JSON de Kev Engine...]<|thought_end|>{  "action": "ast_mutation",  "target_files": ["app/workers/export_worker.py"],  "blast_radius_validated": true,  "target": "finally:\n            pass",  "replacement": "finally:\n            self.semaphore.release()",  "execution_status": "ready"}<|im_end|>

💻 How to Run

1. In Ollama (Recommended)

Download the quantized GGUF (ozygram-qwen7b-Q4_K_M.gguf) and create the model:

bash

ollama create ozygram -f Modelfileollama run ozygram

Modelfile:

dockerfile

FROM ./ozygram-qwen7b-Q4_K_M.gguf PARAMETER temperature 0.2PARAMETER top_p 0.95PARAMETER repeat_penalty 1.1PARAMETER stop "<|im_end|>" TEMPLATE """{{ if .System }}<|im_start|>system{{ .System }}<|im_end|>{{ else }}<|im_start|>system[CONTRATO OZYGRAM NEURAL V1]- Arquitectura: Neuro-Simbólica Local.- Salida del Razonamiento Interno: Entre <|thought_start|> y <|thought_end|>.- Formato de Acción (Kev Engine): JSON estructural estricto sin prosa externa.- Herramientas AST nativas mapeadas en memoria: [Engram_O1_Definitions]<|im_end|>{{ end }}{{ if .Prompt }}<|im_start|>user{{ .Prompt }}<|im_end|>{{ end }}<|im_start|>assistant{{ .Response }}<|im_end|>"""

FROM ./ozygram-qwen7b-Q4_K_M.gguf PARAMETER temperature 0.2PARAMETER top_p 0.95PARAMETER repeat_penalty 1.1PARAMETER stop "<|im_end|>" TEMPLATE """{{ if .System }}<|im_start|>system{{ .System }}<|im_end|>{{ else }}<|im_start|>system[CONTRATO OZYGRAM NEURAL V1]- Arquitectura: Neuro-Simbólica Local.- Salida del Razonamiento Interno: Entre <|thought_start|> y <|thought_end|>.- Formato de Acción (Kev Engine): JSON estructural estricto sin prosa externa.- Herramientas AST nativas mapeadas en memoria: [Engram_O1_Definitions]<|im_end|>{{ end }}{{ if .Prompt }}<|im_start|>user{{ .Prompt }}<|im_end|>{{ end }}<|im_start|>assistant{{ .Response }}<|im_end|>"""

2. In Transformers (Python)

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "Xangel0s/ozyvler-Ozygram-Neural-7B-Coder-MCTS" tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(    model_id,    torch_dtype=torch.bfloat16,    device_map="auto")

⚙️ Hardware Profile & Quantization

Table with columns: Quantization, File Size, GPU VRAM, Recommended Device
Quantization	File Size	GPU VRAM	Recommended Device
`Q4_K_M`	4.36 GB	~4.6 GB	Laptops & Workstations (Quadro RTX 4000, RTX 3060/4060 8GB)
`Q8_0`	~7.5 GB	~8.2 GB	Desktop GPU (12GB+ VRAM)
`BF16`	14.5 GB

📜 Citation

bibtex

@misc{ozygram_neural_v2_2026,  author = {Ozygram Research & Autonomous Engineering Team},  title = {Ozygram Neural v2: Local Neuro-Symbolic MCTS Model for Autonomous Code Self-Healing},  year = {2026},  publisher = {Hugging Face},  howpublished = {\url{https://huggingface.co/Xangel0s/ozyvler-Ozygram-Neural-7B-Coder-MCTS}}}

ozyvler-Ozygram-Neural-7B-Coder-MCTS

README

📌 Model Overview

🏗️ Architecture & Philosophy

📈 Training Metrics & Convergence

📊 Benchmark & Real-World Validation

1. Master Gym Benchmark Suite (100% Pass)

2. Live Local CRM Self-Healing Benchmark (crmgeofa-localdev)

⚡ Prompt Contract & Prefix Caching

💻 How to Run

1. In Ollama (Recommended)

2. In Transformers (Python)

⚙️ Hardware Profile & Quantization

📜 Citation

Explore FriendliAI today

README

📌 Model Overview

🏗️ Architecture & Philosophy

📈 Training Metrics & Convergence

📊 Benchmark & Real-World Validation

1. Master Gym Benchmark Suite (100% Pass)

2. Live Local CRM Self-Healing Benchmark (crmgeofa-localdev)

⚡ Prompt Contract & Prefix Caching

💻 How to Run

1. In Ollama (Recommended)

2. In Transformers (Python)

⚙️ Hardware Profile & Quantization

📜 Citation

2. Live Local CRM Self-Healing Benchmark (`crmgeofa-localdev`)

2. Live Local CRM Self-Healing Benchmark (`crmgeofa-localdev`)