---
license: apache-2.0
base_model: LiquidAI/LFM2.5-1.2B-Instruct
tags:
- linux
- terminal
- bash
- devops
- liquid-foundation-model
- multilingual
- arabic
languages:
- en
- ar
- ta
metrics:
- accuracy
model_name: HydroShell-1.2B
datasets:
- missvector/linux-commands
language:
- en
- ar
---

# HydroShell-1.2B: Liquid Linux Expert
<img src="logo.png" width="50%"/>

**HydroShell-1.2B** is a specialized, multilingual fine-tuned version of the **Liquid AI (LFM 2.5 1.2B)** model. It is optimized to act as a high-performance, low-latency assistant for Linux system administration, shell scripting, and DevOps automation.

By leveraging the **Liquid Foundation Model** architecture, HydroShell excels at processing long-form technical instructions and mapping complex natural language (English, Arabic, and Tamil) to functional Bash one-liners.



## ⚠️ Safety & Destructive Command Warning

> **WARNING:** This model is designed to generate powerful system-level commands. It can and will generate **destructive commands** (e.g., `rm -rf`, `mkfs`, or overwriting configurations with `>`). 
> * **Always verify commands** in a sandbox or test environment before executing them on production systems.
> * The model may occasionally hallucinate flags or mix Linux distributions (e.g., suggesting `pacman` for Ubuntu systems).

---

## Model Details

- **Developed by:** [Your Name/MindLab]
- **Base Model:** LiquidAI/LFM2.5-1.2B-Instruct
- **Architecture:** Liquid Foundation Model (Dynamical Systems-based)
- **Primary Domain:** Linux CLI, Bash Scripting, System Hardening.
- **Languages Supported:** English, Arabic (Technical).



---

## Evaluation Results (Zero-Shot Testing)

The following results were observed during a 100-prompt "Stress Test" covering System Audit, Security, and File Management.

### Technical Performance Matrix
| Category | Accuracy | Notes |
| :--- | :--- | :--- |
| **Basic Admin (`ls`, `cd`, `mkdir`)** | 98% | Flawless execution. |
| **Log Parsing (`awk`, `sed`, `grep`)** | 75% | Occasionally confuses line vs. field flags. |
| **Systemd & Services** | 90% | Strong understanding of service lifecycles. |
| **Networking (`iptables`, `ss`)** | 82% | Occasional source/destination flag inversion. |

### Multilingual Capability
- **Arabic:** 90% Accuracy in intent recognition. Successfully maps Arabic technical terms like "حظر" (Block) and "مزامنة" (Sync).
- **English:** 95% Accuracy in intent recognition.

---

## Known Issues & Limitations
1. **Distro Confusion:** The model may suggest Arch Linux (`pacman`) commands when asked for Ubuntu tasks if the prompt is not specific.
2. **Redirection Risks:** In some tests, the model used `>` (overwrite) instead of `>>` (append) for configuration files. 
3. **Hallucination:** For very complex `find` commands, it may invent non-existent flags (e.g., `-md5`).

---

## Usage (Python)

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "your-username/HydroShell-1.2B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", trust_remote_code=True)

messages = [{"role": "user", "content": "البحث عن العمليات التي تستهلك أكبر قدر من الذاكرة"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=64, temperature=0.3)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

```

## Citation

If you use this model in your research or projects, please cite the base Liquid AI model and this fine-tuned version.

```

---