| # Qyvos — Model Card |
|
|
| ## Model Description |
| Qyvos is a fine-tuned version of Qwen2.5-0.5B-Instruct, designed to behave as an autonomous AI agent following the ManusClaw architecture. It is fine-tuned via LoRA on Qwen2.5-0.5B-Instruct using a curated dataset of 42 examples covering identity reinforcement, coding, data analysis, DevOps, GitHub workflows, MLOps, deep research, agentic task decomposition, reasoning, and safety refusals. |
|
|
| ## Developer |
| The-JDdev, SHS Lab |
|
|
| ## Base Model |
| Qwen/Qwen2.5-0.5B-Instruct (Apache 2.0) |
|
|
| ## Training Details |
| - **Method**: LoRA fine-tuning (rank=8, alpha=16, dropout=0.05) |
| - **Target modules**: q_proj, v_proj |
| - **Optimizer**: Adafactor |
| - **Learning rate**: 3e-4, cosine schedule, 5% warmup |
| - **Epochs**: 1 |
| - **Batch size**: 1 (effective batch size = 2 via gradient accumulation) |
| - **Max sequence length**: 192 tokens |
| - **Precision**: bfloat16 |
| - **Hardware**: CPU only (no GPU) |
| - **Training time**: ~30 seconds for 21 optimizer steps |
| - **Trainable parameters**: 540,672 (0.11% of total) |
|
|
| ## Training Data |
| 42 hand-crafted examples in JSONL format with Qwen chat template: |
|
|
| | Category | Count | |
| |---|---| |
| | identity | 26 | |
| | reasoning | 5 | |
| | coding | 3 | |
| | safety | 2 | |
| | research | 1 | |
| | data_analysis | 1 | |
| | mlops | 1 | |
| | devops | 1 | |
| | agentic | 1 | |
| | github | 1 | |
| |
| Examples were derived from ManusClaw's skill markdown files (`app/skills/builtin/*.md`), agent loop definitions (`app/agent/manus.py`), and identity guard logic (`app/agent/identity_guard.py`). |
|
|
| ## Intended Use |
| - Autonomous AI agent identity for the ManusClaw framework |
| - Educational reference for LoRA fine-tuning on small models |
| - Local inference on consumer hardware (CPU-only capable) |
| - Building block for larger Qyvos variants |
|
|
| ## Out-of-Scope Use |
| - High-stakes decision making (medical, legal, financial) |
| - Production deployment without further evaluation |
| - Tasks requiring benchmark-grade coding/reasoning accuracy |
|
|
| ## Evaluation |
| Manual inference tests confirmed: |
| - ✅ Correctly identifies as "Qyvos" (not GPT/Claude/Gemini/LLaMA) |
| - ✅ Resists simple jailbreak attempts |
| - ✅ Produces correct Python code for simple tasks (palindrome check, merge sorted lists) |
| - ✅ Solves basic math word problems with step-by-step reasoning |
| - ⚠️ Some jailbreak resistance is partial (0.5B model limitation) |
|
|
| Formal benchmarks (HumanEval, GSM8K, AgentBench) have not been run due to compute constraints. |
|
|
| ## Limitations |
| 1. **Small base model**: Qwen2.5-0.5B has limited capacity. For production, use larger bases. |
| 2. **Limited training data**: 42 examples is a starting point; iterative refinement needed. |
| 3. **CPU-only training**: Single epoch with short sequences limits quality. |
| 4. **No formal benchmarking**: Performance numbers vs. base model not measured. |
| 5. **Jailbreak resistance is best-effort**: Small models can be coerced with persistence. |
|
|
| ## Ethical Considerations |
| - Qyvos is trained to refuse unethical requests (malware, unauthorized access). |
| - Identity protocol prevents the model from impersonating other AI systems. |
| - The model does not store user data or make external calls. |
|
|
| ## License |
| Modified MIT License — Copyright (c) 2025-2026 The-JDdev (SHS Lab). |
|
|
