|
|
--- |
|
|
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. |
|
|
|
|
|
``` |
|
|
|
|
|
--- |
