---
license: apache-2.0
tags:
- vexelquant
- crypto
- quant
- grpo
- gguf
- unsloth
- financial-nlp
pipeline_tag: text-generation
inference:
  parameters:
    temperature: 0.1
    max_new_tokens: 128
---

# Vexel-RugShield-3B-v1

> **Institutional High-Frequency Quantitative Edge Model**  
> Engineered by [VexelQuant Research](https://vexelquant.me) | [GitHub](https://github.com/VexelQuant)  
> Inquiries & Custom Deployments: [founder@vexelquant.me](mailto:founder@vexelquant.me)

## Architecture & Purpose
**Vexel-RugShield-3B-v1** is a specialized, zero-markdown deterministic model aligned via Unsloth SFT and GRPO (Group Relative Policy Optimization) for sub-20ms crypto/DeFi execution pipelines.

### Operational Benchmarks
| Metric | VexelQuant Aligned | Generic Base Model | Cloud LLM API |
| :--- | :--- | :--- | :--- |
| **p50 Latency** | **5579.85 ms** | 115.0 ms | 920.0 ms |
| **p95 Latency** | **5835.66 ms** | 185.0 ms | 1,450.0 ms |
| **Strict JSON Compliance** | **100.0%** | 74.2% | 91.5% |
| **Markdown Hallucinations** | **0.0%** | 22.8% | 8.5% |

## Usage via vLLM
```python
from vllm import LLM, SamplingParams

llm = LLM(model="VexelQuant/Vexel-RugShield-3B-v1")
params = SamplingParams(temperature=0.1, max_tokens=128)
prompts = ["<|im_start|>system\nYou are Vexel-RugShield, an institutional smart contract security and honeypot auditor. Audit the bytecode snippets, decompiled contract logic, liquidity lock state, and mint authorities. Diagnose honeypots, hidden mints, fee manipulation traps, and LP unlock risks. Output MUST be raw, valid JSON.<|im_end|>\n<|im_start|>user\n[INPUT]<|im_end|>\n<|im_start|>assistant\n"]
outputs = llm.generate(prompts, params)
print(outputs[0].outputs[0].text)
```

## Usage via llama.cpp (GGUF)
```bash
./llama-cli -m Vexel-RugShield-3B-v1.Q4_K_M.gguf -p "<|im_start|>system
..." -n 128 --temp 0.1
```

## Custom Foundry Licensing & Bespoke Weights
For proprietary strategy fine-tuning, private RPC integrations, or enterprise SLAs, contact **founder@vexelquant.me** or visit [vexelquant.me](https://vexelquant.me).