---

license: apache-2.0

base_model: Qwen/Qwen2.5-Coder-0.5B-Instruct

tags:

• unsloth
• text-generation-inference
• llama-cpp
• gguf
• ollama
• nl2bash
• bash
• terminal-assistant

datasets:

• emirkaanozdemr/bash_command_data_6K

language:

• en

pipeline_tag: text-generation

---

🚀 Qwen2.5-Coder-0.5B-Instruct - NL2Bash (GGUF Quantized)

A fine-tuned version of Qwen/Qwen2.5-Coder-0.5B-Instruct specialized for converting natural language instructions directly into executable Linux/Ubuntu Bash commands with zero conversational filler.

📊 Dual-Track Benchmark Results

🎓 Track 1: Academic NL2Bash Benchmark (EMNLP 2018 / Microsoft CodeXGLUE)

| Metric | Base Model (qwen2.5-coder:0.5b) | Fine-Tuned Model (bash-coder-assistant) | Improvement |

| :--- | :--- | :--- | :--- |

| Exact Match (EM %) | 0.0% | 44.0% | +44.0% |

| Token F1 Score | 12.93 | 81.60 | +68.7 pts |

| BLEU-4 Score | 4.60 | 65.01 | +60.4 pts |

| Bash Syntax Validity | 88.0% | 100.0% | +12.0% |

| Zero-Chatter Clean Format | 4.0% | 100.0% | +96.0% |

| Mean Latency (CPU) | 19.03s | 1.58s | ⚡ 91.7% Faster |

🛠️ Track 2: Real-World DevOps, Cloud & SysAdmin Benchmark

| Metric | Base Model (qwen2.5-coder:0.5b) | Fine-Tuned Model (bash-coder-assistant) | Improvement |

| :--- | :--- | :--- | :--- |

| Exact Match (EM %) | 0.0% | 15.0% | +15.0% |

| Token F1 Score | 6.76 | 62.42 | +55.7 pts |

| Bash Syntax Validity | 90.0% | 100.0% | +10.0% |

| Mean Latency (CPU) | 20.98s | 1.78s | ⚡ 91.5% Faster |

🦙 Quickstart with Ollama

1. Download the qwen2.5-coder-0.5b-bash-gguf.gguf file and Modelfile.
2. Create the model in Ollama:

ollama create bash-coder-assistant -f Modelfile

1. Run a query:

ollama run bash-coder-assistant "Find all processes running on port 8080 and terminate them immediately"
# Output: sudo kill -9 $(lsof -i :8080 | awk '{print $2}')

🛠️ Training Details

• Architecture: Qwen2.5-Coder (0.5B parameters)
• Framework: Unsloth + TRL SFTTrainer
• Dataset: emirkaanozdemr/bash_command_data_6K
• LoRA Rank: r=16, lora_alpha=16
• Training Compute: Google Colab Tesla T4 GPU (< 2GB VRAM, ~8 min runtime)
• Final Loss: 0.6726 (converged from initial 4.1045)