Shell Command Assistant - Gemma 2B LoRA Adapters

This model provides LoRA (Low-Rank Adaptation) adapters for the Gemma 2B model, fine-tuned to help users with shell commands. It can answer questions about common Unix/Linux shell commands with high accuracy.

Model Performance

Accuracy: 100% on core shell commands
Base Model: mlx-community/gemma-2-2b-it-4bit
Adapter Size: ~18MB (much smaller than full model)
Training: Fine-tuned on 26 diverse shell command examples

Quick Start

from mlx_lm import load, generate

# Load the model with adapters
# Step 1: Explicitly download from HF
adapter_path = snapshot_download(repo_id="PKSGIN/MyGemma270_Shellcommands")

# Step 2: Pass the local path to load()
model, tokenizer = load(
    "mlx-community/gemma-2-2b-it-4bit",
    adapter_path=adapter_path  # Now it's a local path!
)

# Ask a question
question = "How do I check disk space?"
prompt = f"### Human: {question}\n### Assistant:"
response = generate(model, tokenizer, prompt=prompt, max_tokens=50, verbose=False)
print(response.strip())

Example Outputs

# Question: How do I list all files including hidden ones?
# Answer: Use ls -la to show all files including hidden ones.

# Question: How do I find all Python files? 
# Answer: Use find . -name "*.py" to find all Python files.

# Question: How do I check memory usage?
# Answer: Use free -h to see memory usage in human-readable format.

Supported Commands

The model can help with these types of shell commands:

File Operations: ls, cp, mv, rm, mkdir, find
System Monitoring: ps, top, free, df
Network: netstat, lsof, ping
File Content: cat, grep, wc, tail, head
Archives: tar, zip, unzip
Permissions: chmod, chown
Process Management: kill, background processes

Training Details

Method: LoRA (Low-Rank Adaptation)
Rank: 4, Alpha: 8, Dropout: 0.0
Training Data: 26 unique shell commands × 60 repetitions = ~1560 examples
Format: Human/Assistant conversation format
Iterations: ~1000 until convergence

Installation

pip install mlx-lm>=0.19.0

Usage Examples

from mlx_lm import load, generate

model, tokenizer = load(
    "mlx-community/gemma-2-2b-it-4bit",
    adapter_path="PKSGIN/MyGemma270_Shellcommands"
)

questions = [
    "How do I compress files?",
    "How do I see running processes?", 
    "How do I find large files?",
    "How do I check network connections?"
]

for question in questions:
    prompt = f"### Human: {question}\n### Assistant:"
    response = generate(model, tokenizer, prompt=prompt, max_tokens=50)
    print(f"Q: {question}")
    print(f"A: {response.strip()}")
    print()

License

Apache 2.0 (following base Gemma model)

Acknowledgments

Built using MLX-LM and based on Google's Gemma 2B model.

PKSGIN
/

MyGemma270_Shellcommands