OrhanDerin/Deepseek-SBX-finetuned

DeepSeek Samba Command Generator (sbx-api)

This repository contains a locally fine-tuned DeepSeek-Coder-1.3b-base model designed to convert natural language requests into precise terminal commands for SAMBA administration.

The model was fine-tuned using LoRA (Low-Rank Adaptation) on a custom dataset of Samba command scenarios. This repository provides both the LoRA adapters and the fully merged standalone model.

Model Details

• Model Type: Causal Language Model (Fine-tuned)
• Base Architecture: deepseek-ai/deepseek-coder-1.3b-base
• Fine-tuning Method: LoRA
• Training Data: Custom Samba command dataset (5,726 samples)
• Language: English
• Intended Use: System administration, specifically generating Samba (samba-tool) commands from text.

Files Provided

This repository contains two sets of model files:

1. Merged Model (deepseek-finetuned_merged/ files): A complete, standalone model (~2.7GB). Ready for direct deployment and use.
2. LoRA Adapter (deepseek-finetuned/ files): Contains only the fine-tuned adapter weights (~250MB). Requires the base model and the peft library to run.

---

How to Get Started with the Model

You can load the model using the Hugging Face transformers library.

Option 1: Using the Merged Model (Recommended)

This is the easiest way to run the model, as it does not require loading separate adapters.

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "OrhanDerin/Deepseek-SBX-finetuned" 

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "Instruction: Create a new user with username jodo and real name john doe\nOutput:"
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(**inputs, max_new_tokens=50)
command = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(command)

Option 2: Using the LoRA Adapter

If you prefer to load the base model and apply the LoRA weights manually:

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model_id = "deepseek-ai/deepseek-coder-1.3b-base"
adapter_id = "OrhanDerin/Deepseek-SBX-finetuned" 

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(base_model_id)

# Load the LoRA adapter onto the base model
model = PeftModel.from_pretrained(base_model, adapter_id)

# Inference works the same as above

---

Training Details

Dataset

The model was trained on a JSON dataset containing 5,726 input-output pairs.

• Input: Natural language command description.
• Output: The exact corresponding Samba command.

Example data format:

{"input": "Create a new user with username jodo and real name john doe", "output": "sudo samba-tool user create jodo --given-name=John --surname=Doe"}
{"input": "enable user kcy", "output": "samba-tool user enable kcy"}
{"input": "get password fields of user kcy", "output": "samba-tool user getpassword kcy"}

Training Hyperparameters

The model was fine-tuned with the following LoRA and training configuration to optimize for 12GB+ VRAM environments:

LoRA Settings:

• Rank (r): 32
• Alpha: 64
• Dropout: 0.05
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, lm_head
• Trainable Parameters: ~21M (1.57% of total 1.34B parameters)

Training Parameters:

• Batch Size: 4 (per device)
• Gradient Accumulation: 4 steps (Effective batch size: 16)
• Learning Rate: 5e-4
• Epochs: 5
• Max Sequence Length: 768 tokens
• Precision: BFloat16
• Optimizer: AdamW with cosine scheduling
• Gradient Checkpointing: Enabled

---

Limitations & Security Notes

Security Warning

• AI-Generated Commands: The commands generated by this model are predictions and should not be executed automatically without human review.
• Destructive Actions: The model may output commands containing sudo, rm -rf, or --force. Always verify the output before running it in a terminal.

Limitations

• Narrow Domain: This model is heavily specialized for Samba (samba-tool) commands. It may hallucinate or perform poorly if asked to generate Python code, general Linux commands outside its training scope, or conversational text.
• Prompt Sensitivity: The model responds best to direct, instructional prompts (e.g., "Instruction: [Task]\nOutput:"). Variations in formatting might yield degraded results.