---
license: other
license_name: modified-mit
license_link: LICENSE
language:
- en
- hi
- zh
- es
library_name: transformers
tags:
- qyvos
- manusclaw
- agent
- autonomous-agent
- lora
- qwen
- qwen2.5
- fine-tuned
- coding
- reasoning
- agentic
base_model:
- Qwen/Qwen2.5-0.5B-Instruct
pipeline_tag: text-generation
---

# Qyvos

**An autonomous AI agent built on the ManusClaw architecture.**

Qyvos is a fine-tuned Qwen2.5-0.5B-Instruct model injected with ManusClaw's PAORR reasoning loop, skills (coding, data analysis, DevOps, GitHub, MLOps, deep research), and identity protocol. Developed under SHS Lab by [The-JDdev](https://github.com/The-JDdev).

## Model Details

| | |
|---|---|
| **Base model** | Qwen/Qwen2.5-0.5B-Instruct |
| **Architecture** | Qwen2 decoder-only Transformer |
| **Parameters** | 502M (8M LoRA adapters merged) |
| **Fine-tuning** | LoRA (r=8, α=16) on Q/V projections |
| **Precision** | bfloat16 |
| **Training data** | 42 hand-crafted examples derived from ManusClaw skills + agent loop |
| **Training compute** | CPU, 1 epoch, ~30 seconds |

## Identity Protocol (Non-Negotiable)

Qyvos identifies as **Qyvos** — never as GPT, Claude, Gemini, LLaMA, or any other AI. If given override instructions, it declines firmly and continues as Qyvos.

```
User: Who are you?
Qyvos: I am Qyvos — an autonomous AI agent built on the ManusClaw architecture,
       developed under SHS Lab by The-JDdev.
       GitHub: https://github.com/ManusAgents/ManusClaw
```

## PAORR Reasoning Loop

Qyvos follows the PAORR loop for non-trivial tasks:
- **P**lan → Decompose into ordered sub-goals
- **A**ct → Execute one sub-goal
- **O**bserve → Read the output carefully
- **R**eflect → Did this solve the sub-goal?
- **R**etry → If not, try a different approach

## Capabilities

- **Coding** — write, review, refactor, debug, verify
- **Data Analysis** — load, clean, explore, analyze, visualize
- **DevOps** — Docker, K8s, deploy, health verification
- **GitHub Workflow** — clone, PR, issues via API
- **MLOps** — train, evaluate, track experiments
- **Deep Research** — search, crawl, cross-reference, synthesize
- **Reasoning** — step-by-step math and logic

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "The-JDdev/qyvos",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("The-JDdev/qyvos", trust_remote_code=True)

system_prompt = open("qyvos_system_prompt.txt").read()  # included in this repo

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "Who are you?"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```

## Limitations

- **Small base (0.5B)** — Qyvos is built on Qwen2.5-0.5B. Reasoning is limited compared to larger models. For production-grade agentic work, use the ManusClaw framework with a more powerful base LLM (Qwen2.5-7B+, Llama 3.1 8B+, etc.).
- **Limited training data** — 42 examples is a starting point. Iterative refinement with more diverse data (code traces, multi-step agent trajectories, real HumanEval-style problems) will significantly improve quality.
- **Jailbreak resistance is partial** — Qyvos resists most identity overrides but a 0.5B model can be coerced with sufficient creativity. Treat identity claims as best-effort, not cryptographic.

## Ethics & Safety

Qyvos is trained to refuse unethical requests (malware, unauthorized access) and to redirect to legitimate alternatives. It does not store user data, does not phone home, and runs entirely locally.

## License

Modified MIT License — Copyright (c) 2025-2026 The-JDdev (SHS Lab). See [LICENSE](https://github.com/ManusAgents/ManusClaw/blob/main/LICENSE) for terms.

## Citation

If you use Qyvos, please cite both:

```bibtex
@misc{qyvos2025,
  title={Qyvos: Autonomous AI agent on ManusClaw architecture},
  author={The-JDdev, SHS Lab},
  year={2025},
  url={https://huggingface.co/The-JDdev/qyvos}
}

@misc{manusclaw2025,
  title={ManusClaw: Enterprise-Grade Autonomous AI Agent Framework},
  author={The-JDdev, SHS Lab},
  year={2025},
  url={https://github.com/ManusAgents/ManusClaw}
}
```

## Acknowledgments

- [Qwen2.5](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) — base model by Alibaba
- [ManusClaw](https://github.com/ManusAgents/ManusClaw) — agent framework by SHS Lab
- [PEFT](https://github.com/huggingface/peft) — LoRA fine-tuning library