README.md

metadata

title: Fistal AI
emoji: 🚀
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.0.1
app_file: app.py
pinned: false
license: apache-2.0
short_description: Finetuning Studio
python_version: 3.11
tags:
  - mcp-in-action-track-enterprise
  - mcp-in-action-track-consumer

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

🚀 Fistal AI - Autonomous Fine-Tuning Platform

Agentic AI that seamlessly finetunes LLM's with Unsloth and Modal

🎮 Try Demo • 📱 LinkedIn Post

---

🎯 What is Fistal AI?

Fistal AI is an autonomous fine-tuning platform that transforms the complex process of training custom language models into a single-click experience. Simply specify your topic, and Fistal handles everything:

• 🤖 Synthetic Dataset Generation - Creates high-quality training data using LLMs
• 🔄 Automatic Data Formatting - Converts to chat/instruction format
• 🏋️ Serverless Training - Fine-tunes models on Modal's GPU infrastructure
• 📊 LLM-as-Judge Evaluation - Validates model performance
• 🤗 Hugging Face Deployment - Publishes your model automatically

No ML expertise required. No infrastructure setup. Just results.

---

✨ Features

🎨 Intuitive Interface

• Clean Gradio-based web UI hosted on Hugging Face Spaces
• Real-time training progress with educational insights
• Automatic Hugging Face integration with one-click model access
• Direct model upload in native HF format (ready to use immediately)

⚡ Blazing Fast Training

• 3x faster dataset generation with parallel API calls (Gemini)
• 2x faster training with Unsloth optimization and Modal GPU's
• 70% less memory usage via 4-bit quantization
• Training completes in 10-20 minutes for 500 samples

🧠 Smart Defaults

• 4-bit quantization for optimal quality/size balance
• LoRA fine-tuning (updates only 0.1% of parameters)
• Supports 1B-3B parameter models (Qwen, Llama, Gemma, Phi)
• Automatic hyperparameter optimization
• Native HF format upload (no conversion needed)

🔬 Quality Assurance

• LLM-as-judge evaluation system
• Coherence, relevance, and accuracy testing
• Comprehensive evaluation reports
• Real-time monitoring of training metrics

🔌 MCP-Powered Workflow

• Agentic orchestration using Model Context Protocol (MCP)
• 4 specialized MCP tools for end-to-end automation
• Intelligent decision-making throughout the pipeline
• Seamless tool coordination for optimal results

---

✨ Sponsors:

• Modal Labs : Seamless T4 GPU access
• Gemini API : Handles majority of LLM tasks including data generation and agentic control

---

Watch Demo:

Note: The demo runs only 5 samples for speed, but you can scale it to 2000+ in real use.

Demo Fistal

---

How Fistal AI Works Behind the Scenes

• Fistal AI runs on an agentic workflow powered by LangGraph.
• Instead of a fixed script, an AI agent decides what step to run next.
• All the actual work (dataset generation, formatting, training, evaluation) is done by MCP tools.
• The agent just thinks → MCP tools do the work → agent continues automatically.

---

The MCP Server

• Hosts four tools:

  • generate_json_data → creates synthetic training data
  • format_json → converts it to ChatML format
  • finetune_model → runs Unsloth training on Modal
  • llm_as_judge → evaluates the trained model

• Each tool is isolated and safe.

• Returns clean, structured results that the agent uses.

---

Pipeline Flow (Step-by-Step)

• 1. Dataset Generation Agent calls the tool → LLMs generate 20–500 examples in parallel.

• 2. Dataset Formatting Agent calls next tool → raw dataset becomes ChatML/instruction format.

• 3. Fine-Tuning Agent launches training on Modal using Unsloth + 4-bit QLoRA.

• 4. Evaluation Agent runs LLM-as-judge → gets coherence/relevance/accuracy/ROUGE/BLEU scores with evaluate library.

• 5. Final Output The model and adapters are automatically uploaded to the user's(mahreenfathima) Hugging Face account (based on the HF token provided). Automatic Evaluation Report generated.

---

🛠️ The 4 MCP Tools

1. generate_json_data

   • Purpose: Synthetic dataset generation
   • Input: Topic, sample count, task type
   • Process: Parallel API calls to Gemini + Groq with intelligent prompt engineering
   • Output: JSON dataset with diverse, high-quality examples
   • MCP Role: Agent invokes this tool first, receives confirmation, then proceeds

2. format_json

   • Purpose: Convert raw data to training(ChatML) format
   • Input: Raw JSON dataset path
   • Process: Transforms to chat/instruction format optimized for fine-tuning
   • Output: Formatted dataset ready for training
   • MCP Role: Agent receives dataset path from previous tool, formats it automatically

3. finetune_model

   • Purpose: Execute serverless training
   • Input: Formatted dataset, model name, hyperparameters
   • Process: Deploys training job to Modal with Unsloth optimization
   • Output: Fine-tuned model weights + training metrics
   • MCP Role: Agent monitors training progress, handles failures, manages GPU resources
   • Internal Functions (executed within Modal):
     • train_with_modal: Runs finetuning process with Unsloth and saves model in Volume
     • upload_to_hf_from_volume: Pushes the trained model weights to Hugging Face Hub repository

4. llm_as_judge

   • Purpose: Quality evaluation
   • Input: Fine-tuned model path, test cases
   • Process: Generates test prompts, evaluates responses, scores quality
   • Output: Comprehensive evaluation report with metrics
   • MCP Role: Final validation step, agent parses results and presents to user

• Internal Functions (executed within Modal):
  • evaluate_model: Runs validation metrics on the fine-tuned model during/after training

🧠 Fistal's Agentic Approach

# Agent makes decisions based on context
agent decides: "User wants Python dataset"
  → invokes generate_json_data with optimal parameters
  
agent observes: "Dataset generated successfully"
  → invokes format_json with received path
  
agent monitors: "Training at 50%, loss decreasing"
  → continues monitoring, adjusts if needed
  
agent validates: "Model trained, run evaluation"
  → invokes llm_as_judge for quality check

Benefits:

• 🎯 Intelligent Decision Making: Agent chooses best parameters and strategies
• 🔄 Error Recovery: Automatically retries failed steps with adjusted parameters
• 📊 Context Awareness: Each tool receives relevant context from previous steps
• 🔒 Security: MCP provides secure tool execution
• 🔧 Modularity: Tools can be updated independently without breaking the workflow
• 📈 Scalability: Easy to add new tools (e.g., hyperparameter tuning, multi-GPU training)

---

🛠️ Tech Stack

Core Technologies Unsloth - 2x faster training, 70% less VRAM Modal - Serverless GPU infrastructure Gradio - Web interface on HF Spaces LangGraph - Agentic workflow orchestration MCP - Tool integration protocol HUGGING FACE - Uploads model into repository with hf tokens

AI Models & APIs Gemini Flash 2.0 - Fast dataset generation Groq (Llama 3.1 70B) - LLM evaluation Hugging Face - Model hosting & deployment 4-bit Quantization - Optimal quality/size balance Native HF Upload - No format conversion needed

---

📊 Performance Metrics

Metric	Value	Details
Dataset Generation	3x faster	Parallel processing with API keys
Training Speed	2x faster	Unsloth optimization
Memory Usage	-70%	4-bit quantization
Training Time	10-20 min	For 500 samples on T4 GPU
Model Size	~1-2 GB	Native HF format (safetensors)
Parameters Updated	0.1%	LoRA efficiency
MCP Tools	4	Autonomous workflow management

---

🔧 Supported Models & Tasks

Prominent Models (1B-3B Parameters)

• Qwen/Qwen2.5-1.5B-Instruct
• Qwen/Qwen2.5-3B-Instruct
• meta-llama/Llama-3.2-1B-Instruct
• meta-llama/Llama-3.2-3B-Instruct
• google/gemma-2-2b-it
• microsoft/Phi-3.5-mini-instruct

Popular Task Types

• text-generation: General text completion and content creation
• question-answering: Q&A pairs and knowledge retrieval

Output Format

• Native Hugging Face format (safetensors + adapter weights)
• Immediately usable with transformers library
• Compatible with HF Inference API

---

🎮 Try It Now

🚀 Launch Fistal AI Demo

📱 Read LinkedIn Post

Hosted on Hugging Face Spaces - No installation required!

---

📝 License

This project is licensed under the APACHE License - see the LICENSE file for details.

---

🙏 Acknowledgments

• Anthropic MCP - For the powerful tool integration protocol
• Unsloth - For making fine-tuning accessible and fast
• Modal - For serverless GPU infrastructure
• Hugging Face - For model hosting and Spaces platform
• Google Gemini - For powerful API access
• LangGraph - For agentic orchestration framework
• Gradio - For building the interactive UI effortlessly

---

Powered by MCP • Unsloth • Modal • Hugging Face • Gemini API

❤️ Like our space our HuggingFace • 🚀 Try the demo • 📱 Share on LinkedIn