tunedai/myguardian-guidance

MyGuardian Guidance AI — HF Spaces Deployment

Fine-tuned causal reasoning model for real-time police interaction legal guidance.

Project Structure

myguardian-hf-spaces/
├── app.py                          # Gradio app for inference
├── train_data.jsonl                # Training examples (24)
├── eval_data.jsonl                 # Eval examples (7)
├── prepare_finetuning_dataset.py   # Dataset generation
├── requirements.txt
└── README.md

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Run Locally (with base model)

python3 app.py

Open http://localhost:7860 in your browser.

Fine-tuning

Option A: Use HF AutoTrain (Recommended for limited budget)

1. 1.Upload training data to HF Hub:

   huggingface-cli repo create myguardian-guidance --type dataset
   huggingface-cli upload huggingface_username/myguardian-guidance \
     train_data.jsonl eval_data.jsonl .

1. 1.Create AutoTrain project:
2. 2.Go to https://huggingface.co/autotrain
3. 3.Create new project → "Language Model Fine-tuning"
4. 4.Select dataset: huggingface_username/myguardian-guidance
5. 5.Base model: Qwen/Qwen2.5-7B
6. 6.Training params:
7. 7.Learning rate: 1e-4
8. 8.Num epochs: 3
9. 9.Batch size: 4
10. 10.LoRA (for memory efficiency)
11. 11.Start training

1. 1.Once complete, get the model URL from AutoTrain and update app.py:

   MODEL_NAME = "huggingface_username/myguardian-guidance"

Option B: Local Fine-tuning (for testing)

from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from datasets import load_dataset

# Load data
dataset = load_dataset('json', data_files={
    'train': 'train_data.jsonl',
    'eval': 'eval_data.jsonl'
})

# Load model
model_name = "Qwen/Qwen2.5-7B"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Training arguments
args = TrainingArguments(
    output_dir="./myguardian-guidance",
    num_train_epochs=3,
    learning_rate=1e-4,
    per_device_train_batch_size=4,
    per_device_eval_batch_size=4,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    logging_steps=10,
)

# Train
trainer = Trainer(
    model=model,
    args=args,
    train_dataset=dataset['train'],
    eval_dataset=dataset['eval'],
)
trainer.train()

Deployment to HF Spaces

1. Create HF Spaces Repo

huggingface-cli repo create myguardian-guidance --type space --space-sdk gradio
cd ~/my-spaces-clones/myguardian-guidance

2. Push Code

git add app.py requirements.txt
git commit -m "Initial commit"
git push

HF Spaces will automatically build and run the app.

3. Update Model Name

In app.py, change:

MODEL_NAME = "huggingface_username/myguardian-guidance"

The model will be downloaded on first load.

Cost Estimates

• —Local fine-tuning: $0 (uses your machine)
• —HF AutoTrain (8B model, 3 epochs): ~$30-50
• —Inference on HF Spaces (free tier): Free (respects rate limits)

Total budget: $40-80 ✓

Test Harness

The 31 test scenarios achieve 87% accuracy with the original guardian_api.py (using GPT-4).

To evaluate the fine-tuned model, compare its output against test scenarios in:

~/Desktop/myguardian_demo/test_harness.py

Expected outputs should match:

• —STATE: Correct situation classification
• —SAY: Actionable, jurisdiction-specific advice
• —AVOID: Common legal mistakes
• —OPTIONS: Causal reasoning (if X, then Y)

Notes

• —Model: Qwen2.5-7B (7B parameters, fast, good for mobile)
• —Training data: 31 real police interaction scenarios
• —Temperature: 0.3 (deterministic for safety)
• —Max tokens: 500 (keeps responses concise)
• —No hallucination penalties — focus on causal accuracy

Next Steps

1. 1.✓ Prepare training data from test scenarios
2. 2.⚠️ Fine-tune on HF or locally
3. 3.⚠️ Evaluate against test harness
4. 4.⚠️ Deploy to HF Spaces
5. 5.⚠️ Share URL with CEO for feedback

License

Internal TunedAI Labs / MyGuardian. See licensing agreement.