README.md

---
language:
- ru
- en
license: apache-2.0
tags:
- linux
- shell
- qwen2
- unsloth
- gguf
base_model: Qwen/Qwen2.5-3B-Instruct
pipeline_tag: text-generation
---


# Qwen2.5-3B Linux Assistant

A fine-tuned version of [Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) trained to act as a Linux/Shell command assistant. Given a natural language description, the model outputs the correct shell command.

Supports both **Russian** and **English** input.

---

## Model Details

| Property | Value |
|---|---|
| Base model | Qwen2.5-3B-Instruct |
| Fine-tuning method | QLoRA (LoRA r=16, alpha=16) |
| Training steps | ~1700 |
| Epochs | 3 |
| Final loss | ~0.28 |
| Dataset size | ~4500 examples |
| Languages | Russian, English |
| Framework | Unsloth + TRL |

---

## Usage

### Ollama (recommended)
```bash
ollama run hf.co/NickIBrody/qwen-linux-gguf
```

### llama.cpp
```bash
llama-cli -hf NickIBrody/qwen-linux-gguf --jinja
```

### Python (transformers)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "NickIBrody/qwen-linux"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")

messages = [
    {"role": "system", "content": "You are a Linux assistant. Reply only with the shell command, no explanations."},
    {"role": "user", "content": "show top 5 processes by memory usage"},
]
inp = tok.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inp, max_new_tokens=128, temperature=0.3)
print(tok.decode(out[0][inp.shape[1]:], skip_special_tokens=True))
```

---

## Examples

| Input | Output |
|---|---|
| показажи топ 5 процессов по памяти | `ps aux --sort=-%mem \| head -n 5` |
| где я нахожусь в терминале | `pwd` |
| compress file data.txt with bzip2 | `bzip2 data.txt` |
| show disk usage in human readable format | `df -h` |
| find all .log files modified in last 7 days | `find / -name "*.log" -mtime -7` |
| kill process by name nginx | `pkill nginx` |
| show open ports | `ss -tulnp` |

---

## Dataset

Training data: [NickIBrody/linux-commands-ru-en](https://huggingface.co/datasets/NickIBrody/linux-commands-ru-en)

~4500 shell command examples in Russian and English, covering:
- File system navigation and management
- Process management
- Networking
- Archive and compression
- System monitoring
- Package management

---

## Training Code

```python
from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template
from datasets import load_dataset
from trl import SFTTrainer
from transformers import TrainingArguments

model, tok = FastLanguageModel.from_pretrained(
    "unsloth/Qwen2.5-3B-Instruct-bnb-4bit",
    max_seq_length=2048,
    load_in_4bit=True
)

model = FastLanguageModel.get_peft_model(
    model, r=16, lora_alpha=16,
    target_modules=["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"]
)

tok = get_chat_template(tok, chat_template="qwen-2.5")

ds = load_dataset("NickIBrody/linux-commands-ru-en", split="train")
ds = ds.map(lambda x: {"text": tok.apply_chat_template(x["messages"], tokenize=False)})

SFTTrainer(
    model=model,
    tokenizer=tok,
    train_dataset=ds,
    dataset_text_field="text",
    max_seq_length=2048,
    args=TrainingArguments(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        num_train_epochs=3,
        learning_rate=2e-4,
        fp16=True,
        logging_steps=10,
        output_dir="out",
        optim="adamw_8bit"
    )
).train()

model.save_pretrained_gguf("qwen-linux", tok, quantization_method="q4_k_m")
```

---

## Limitations

- Designed for shell commands only, not general conversation
- May struggle with highly complex multi-step scripts
- Best results with clear, specific prompts

---

## License

Apache 2.0