🔧 CLI LoRA TinyLLaMA Fine-Tuning (Fenrir Internship Project)

🚀 This repository presents a LoRA fine-tuned version of TinyLLaMA-1.1B-Chat trained on a custom dataset of CLI Q&A. Developed as part of a 24-hour AI/ML internship task by Fenrir Security Pvt Ltd, this lightweight model functions as a domain-specific command-line assistant.

📁 Dataset

A curated collection of 200+ real-world CLI Q&A pairs covering:

Git (branching, stash, merge, rebase)
Bash (variables, loops, file manipulation)
grep, tar, gzip (command syntax, flags)
Python environments (venv, pip)

Stored in cli_questions.json.

⚙️ Model Details

Field	Value
Base Model	`TinyLlama/TinyLlama-1.1B-Chat-v1.0`
Fine-Tuning Method	QLoRA via `peft`
Epochs	3 (with early stopping)
Adapter Size	~7MB (LoRA weights only)
Hardware	Local CPU (low-resource)
Tokenizer	Inherited from base model

📊 Evaluation

Metric	Result
Accuracy on Eval Set	~92%
Manual Review	High relevance
Hallucination Rate	Very low
Inference Time (CPU)	< 1s / query

All results are stored in eval_results.json.

🧠 Files Included

adapter_model.safetensors — fine-tuned LoRA weights
adapter_config.json — LoRA hyperparameters
training.ipynb — complete training notebook
agent.py — CLI interface to test the model
cli_questions.json — training dataset
eval_results.json — eval results
requirements.txt — dependencies

📦 Inference Example

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0")
tokenizer = AutoTokenizer.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0")

peft_model = PeftModel.from_pretrained(base_model, "Harish2002/cli-lora-tinyllama")
peft_model.eval()

prompt = "How do I initialize a new Git repository?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = peft_model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))