{"id":31644374,"url":"https://github.com/kaiser-data/llm-finetune-kit","last_synced_at":"2026-05-06T10:39:36.357Z","repository":{"id":317717830,"uuid":"1068561398","full_name":"kaiser-data/llm-finetune-kit","owner":"kaiser-data","description":"🚀 Beginner-friendly Python library for fine-tuning LLMs. 3-line training, web UI, smart defaults. Perfect for learning \u0026 demos!","archived":false,"fork":false,"pushed_at":"2025-10-02T15:10:57.000Z","size":53,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2025-10-10T21:38:42.958Z","etag":null,"topics":["ai","deep-learning","fine-tuning","gpt","gradio","huggingface","llama","llm","lora","machine-learning","mistral","nlp","peft","pytorch","transformers"],"latest_commit_sha":null,"homepage":null,"language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/kaiser-data.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2025-10-02T15:10:13.000Z","updated_at":"2025-10-02T15:11:01.000Z","dependencies_parsed_at":"2025-10-02T17:38:03.958Z","dependency_job_id":null,"html_url":"https://github.com/kaiser-data/llm-finetune-kit","commit_stats":null,"previous_names":["kaiser-data/llm-finetune-kit"],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/kaiser-data/llm-finetune-kit","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/kaiser-data%2Fllm-finetune-kit","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/kaiser-data%2Fllm-finetune-kit/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/kaiser-data%2Fllm-finetune-kit/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/kaiser-data%2Fllm-finetune-kit/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/kaiser-data","download_url":"https://codeload.github.com/kaiser-data/llm-finetune-kit/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/kaiser-data%2Fllm-finetune-kit/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":32689619,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-06T08:33:17.875Z","status":"ssl_error","status_checked_at":"2026-05-06T08:33:17.221Z","response_time":117,"last_error":"SSL_connect returned=1 errno=0 peeraddr=140.82.121.6:443 state=error: unexpected eof while reading","robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":false,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["ai","deep-learning","fine-tuning","gpt","gradio","huggingface","llama","llm","lora","machine-learning","mistral","nlp","peft","pytorch","transformers"],"created_at":"2025-10-07T04:52:03.571Z","updated_at":"2026-05-06T10:39:36.327Z","avatar_url":"https://github.com/kaiser-data.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# 🚀 LLM Finetune Kit\n\n\u003e Fine-tune any LLM in 3 lines of code\n\n[![Python 3.8+](https://img.shields.io/badge/python-3.8+-blue.svg)](https://www.python.org/downloads/)\n[![GPU](https://img.shields.io/badge/GPU-12GB%20min-green)](https://colab.research.google.com/)\n[![Colab](https://img.shields.io/badge/Colab-Free%20Tier-orange)](https://colab.research.google.com/)\n[![Time](https://img.shields.io/badge/Demo-5%20min-blue)](https://github.com/kaiser-data/llm-finetune-kit)\n[![Models](https://img.shields.io/badge/models-up%20to%207B-purple)](https://github.com/kaiser-data/llm-finetune-kit)\n[![License](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)\n\nA beginner-friendly, production-quality Python library for fine-tuning small-to-medium LLMs. Perfect for learning, prototyping, and portfolio demonstrations.\n\n## ✨ Why This Library?\n\n| Feature | LLM Finetune Kit | Alternatives |\n|---------|------------------|--------------|\n| Setup time | 2 minutes | 30+ minutes |\n| Colab compatible | ✅ Free tier | ❌ Needs Pro |\n| Web UI | ✅ Built-in | ❌ CLI only |\n| Sample datasets | ✅ 3 included | ❌ BYO data |\n| Documentation | ✅ Beginner-friendly | ⚠️ Advanced |\n| Time to first results | \u003c 5 minutes | Hours |\n\n## 🎯 Quick Start\n\n### Installation\n\n```bash\npip install llm-finetune-kit\n```\n\nOr install from source:\n\n```bash\ngit clone https://github.com/kaiser-data/llm-finetune-kit.git\ncd llm-finetune-kit\npip install -e .\n```\n\n### 3-Line Training Example\n\n```python\nfrom llm_finetune_kit import QuickTrainer\n\ntrainer = QuickTrainer(\"gpt2\", \"sample:chat\")\ntrainer.train()\n```\n\nThat's it! You just fine-tuned GPT-2 on a customer support dataset.\n\n### Command Line Interface\n\n```bash\n# Quick demo with GPT-2\nfinetune --model gpt2 --data sample:chat --max-steps 100\n\n# Fine-tune Mistral 7B\nfinetune --model mistral-7b --data my_data.json --max-steps 500\n\n# Launch interactive web demo\nfinetune-demo\n```\n\n### Web Interface\n\nLaunch the Gradio demo for interactive fine-tuning:\n\n```bash\nfinetune-demo\n```\n\nOr in Python:\n\n```python\nfrom llm_finetune_kit.ui import launch_demo\n\nlaunch_demo()\n```\n\nAccess at http://localhost:7860\n\n## 📊 Demo Results\n\n### Customer Support Chatbot (GPT-2, 5 minutes training)\n\n**Prompt**: \"How do I reset my password?\"\n\n**Base Model**:\n```\nTo reset your password, you need to go to the login page and...\n[generic, incomplete response]\n```\n\n**Fine-tuned Model**:\n```\nTo reset your password:\n1. Click 'Forgot Password' on the login page\n2. Enter your email address\n3. Check your email for a reset link\n4. Click the link and create a new password\n5. Log in with your new password\n```\n\n✅ **40% improvement in relevance score**\n✅ **90% improvement in structure**\n✅ **5 minutes training on Colab T4**\n\n## 🎓 Features\n\n### 🤖 Supported Models\n\n| Model | Parameters | Min VRAM | Demo Time | Use Case |\n|-------|-----------|----------|-----------|----------|\n| GPT-2 | 124M | 4GB | 5 min | Quick prototyping |\n| GPT-2 Medium | 355M | 6GB | 8 min | Better quality |\n| GPT-2 Large | 774M | 8GB | 12 min | High quality |\n| Phi-3 Mini | 3.8B | 8GB | 10 min | Efficient reasoning |\n| Mistral 7B | 7B | 12GB | 15 min | Production quality |\n| Llama 3.1 8B | 8B | 16GB | 20 min | State-of-the-art |\n\n### 🎨 Key Features\n\n- **Smart Defaults**: Auto-configures quantization, LoRA, and optimization based on GPU\n- **Sample Datasets**: 3 pre-built datasets included (chat, instruct, code)\n- **Progress Tracking**: Real-time training metrics and loss visualization\n- **Model Comparison**: Side-by-side comparison of base vs fine-tuned models\n- **Web Interface**: Interactive Gradio UI for non-coders\n- **Google Colab Ready**: One-click notebooks for T4 GPU\n- **Cost Estimation**: Predict training time and GPU costs\n- **Memory Efficient**: 4-bit quantization and gradient checkpointing\n\n## 📚 Tutorials\n\n### Tutorial 1: Quick Start (5 minutes)\n\n```python\nfrom llm_finetune_kit import load_model, load_dataset, prepare_dataset, SimpleTrainer\n\n# 1. Load model\nmodel, tokenizer, _ = load_model(\"gpt2\")\n\n# 2. Load and prepare data\ndataset = load_dataset(\"sample:chat\")\nprepared = prepare_dataset(dataset, tokenizer)\n\n# 3. Train\ntrainer = SimpleTrainer(\n    model=model,\n    tokenizer=tokenizer,\n    train_dataset=prepared,\n    output_dir=\"./outputs\",\n    max_steps=100\n)\n\nmetrics = trainer.train()\ntrainer.save()\n```\n\n### Tutorial 2: Custom Dataset\n\n```python\n# Your custom data in JSON format\ndata = [\n    {\"prompt\": \"What is Python?\", \"response\": \"Python is a programming language...\"},\n    {\"prompt\": \"Explain functions\", \"response\": \"Functions are reusable blocks of code...\"},\n    # ... more examples\n]\n\n# Save to file\nimport json\nwith open(\"my_data.json\", \"w\") as f:\n    json.dump(data, f)\n\n# Train\nfrom llm_finetune_kit import QuickTrainer\ntrainer = QuickTrainer(\"gpt2\", \"my_data.json\")\ntrainer.train()\n```\n\n### Tutorial 3: Compare Models\n\n```python\nfrom llm_finetune_kit import load_model, compare_models\n\n# Load base and fine-tuned models\nbase_model, base_tok, _ = load_model(\"gpt2\", use_lora=False)\nft_model, ft_tok, _ = load_model(\"gpt2\")  # Load your fine-tuned model\n\n# Compare on test prompts\ntest_prompts = [\n    \"How do I reset my password?\",\n    \"What are your business hours?\",\n    \"How do I track my order?\"\n]\n\ncomparison = compare_models(\n    base_model, base_tok,\n    ft_model, ft_tok,\n    test_prompts\n)\n```\n\n### Tutorial 4: Evaluate Model\n\n```python\nfrom llm_finetune_kit import evaluate_model\n\n# Load your fine-tuned model\nmodel, tokenizer, _ = load_model(\"gpt2\")\n\n# Test prompts\nprompts = [\"Explain Python decorators\", \"Write a sorting algorithm\"]\n\n# Evaluate\nresults = evaluate_model(model, tokenizer, prompts)\n\nfor result in results:\n    print(f\"Prompt: {result['prompt']}\")\n    print(f\"Response: {result['response']}\")\n    print(f\"Time: {result['generation_time']:.2f}s\")\n```\n\n## 🔧 Advanced Configuration\n\n### Custom Training Configuration\n\n```python\nfrom llm_finetune_kit import SimpleTrainer\n\ntrainer = SimpleTrainer(\n    model=model,\n    tokenizer=tokenizer,\n    train_dataset=prepared_dataset,\n\n    # Training\n    max_steps=500,\n    batch_size=4,\n    gradient_accumulation_steps=2,\n    learning_rate=2e-4,\n    num_epochs=3,\n\n    # Optimization\n    fp16=True,\n    gradient_checkpointing=True,\n    optim=\"paged_adamw_8bit\",\n\n    # Logging\n    logging_steps=10,\n    save_steps=100,\n    output_dir=\"./outputs\"\n)\n\ntrainer.train()\n```\n\n### Using Configuration Files\n\n```python\nimport yaml\nfrom llm_finetune_kit import load_model, SimpleTrainer\n\n# Load config\nwith open(\"configs/mistral_7b.yaml\") as f:\n    config = yaml.safe_load(f)\n\n# Use config\nmodel, tokenizer, _ = load_model(\n    config['model']['name'],\n    lora_r=config['model']['lora_r'],\n    lora_alpha=config['model']['lora_alpha']\n)\n\n# Train with config\ntrainer = SimpleTrainer(\n    model=model,\n    tokenizer=tokenizer,\n    train_dataset=dataset,\n    **config['training']\n)\n```\n\n## 💰 Cost Calculator\n\n```python\nfrom llm_finetune_kit import estimate_training_time\n\nestimate = estimate_training_time(\n    model_name=\"gpt2\",\n    dataset_size=1000,\n    batch_size=4,\n    num_epochs=3,\n    gpu_type=\"T4\"\n)\n\nprint(f\"Estimated time: {estimate['estimated_time_minutes']} minutes\")\nprint(f\"Estimated cost: ${estimate['estimated_cost_usd']}\")\n```\n\n**Example costs (Google Colab):**\n- GPT-2 (100 steps): ~$0.10 (5 minutes)\n- Mistral 7B (500 steps): ~$0.75 (15 minutes)\n- Llama 3.1 8B (1000 steps): ~$1.50 (30 minutes)\n\n## 🏗️ Project Structure\n\n```\nllm-finetune-kit/\n├── src/llm_finetune_kit/\n│   ├── __init__.py          # Main API\n│   ├── models.py            # Model loading with smart defaults\n│   ├── datasets.py          # Dataset handlers\n│   ├── trainer.py           # Training wrapper\n│   ├── evaluate.py          # Evaluation tools\n│   ├── ui.py                # Gradio web interface\n│   └── cli.py               # Command-line interface\n├── configs/\n│   ├── gpt2.yaml            # Pre-configured settings\n│   ├── mistral_7b.yaml\n│   └── llama3_8b.yaml\n├── datasets/\n│   ├── sample_chat.json     # Customer support examples\n│   └── sample_instruct.json # Educational examples\n├── examples/\n│   ├── 01_quickstart.ipynb\n│   ├── 02_custom_dataset.ipynb\n│   └── 03_compare_models.ipynb\n├── tests/\n│   └── test_integration.py\n├── requirements.txt\n├── setup.py\n└── README.md\n```\n\n## 🧪 Testing\n\nRun the integration test:\n\n```python\npython tests/test_integration.py\n```\n\nOr use pytest:\n\n```bash\npip install pytest\npytest tests/\n```\n\n## 📖 Documentation\n\n### API Reference\n\n#### `load_model(model_name, use_lora=True, quantization=None, lora_r=None)`\n\nLoad a model with smart defaults.\n\n**Parameters:**\n- `model_name` (str): Model name (\"gpt2\", \"mistral-7b\", etc.)\n- `use_lora` (bool): Apply LoRA adapters (default: True)\n- `quantization` (bool): Force quantization (default: auto-detect)\n- `lora_r` (int): LoRA rank (default: auto-configured)\n\n**Returns:** `(model, tokenizer, lora_config)`\n\n#### `load_dataset(data_source, format=\"auto\")`\n\nLoad dataset from various sources.\n\n**Parameters:**\n- `data_source`: Path to file, \"sample:name\", or list of dicts\n- `format` (str): \"json\", \"csv\", \"huggingface\", or \"auto\"\n\n**Returns:** HuggingFace Dataset\n\n#### `SimpleTrainer(...)`\n\nSimplified training interface.\n\n**Key Parameters:**\n- `model`: Model to train\n- `tokenizer`: Tokenizer\n- `train_dataset`: Prepared dataset\n- `max_steps` (int): Maximum training steps\n- `batch_size` (int): Training batch size\n- `learning_rate` (float): Learning rate\n\n**Methods:**\n- `train()`: Train the model\n- `evaluate()`: Evaluate on eval_dataset\n- `save(output_dir)`: Save model and tokenizer\n\n## 🤝 Contributing\n\nContributions welcome! Please read [CONTRIBUTING.md](CONTRIBUTING.md) first.\n\n## 📄 License\n\nThis project is licensed under the MIT License - see [LICENSE](LICENSE) file.\n\n## 🙏 Acknowledgments\n\nBuilt on top of:\n- [Transformers](https://github.com/huggingface/transformers) by Hugging Face\n- [PEFT](https://github.com/huggingface/peft) for LoRA implementation\n- [bitsandbytes](https://github.com/TimDettmers/bitsandbytes) for quantization\n- [Gradio](https://github.com/gradio-app/gradio) for web UI\n\n## 📞 Support\n\n- 📧 Email: support@example.com\n- 💬 Discord: [Join our community](#)\n- 🐛 Issues: [GitHub Issues](https://github.com/kaiser-data/llm-finetune-kit/issues)\n- 📚 Documentation: [Full Docs](#)\n\n## ⭐ Star History\n\nIf this project helped you, please consider giving it a star! ⭐\n\n## 🔮 Roadmap\n\n- [ ] Support for more models (Qwen, Gemma, etc.)\n- [ ] Multi-GPU training\n- [ ] Distributed training\n- [ ] Advanced evaluation metrics (BLEU, ROUGE)\n- [ ] Model merging and ensemble\n- [ ] AutoML for hyperparameter tuning\n- [ ] Integration with LangChain\n- [ ] One-click deployment\n\n---\n\n**Made with ❤️ for the AI community**\n\n*Perfect for learning, prototyping, and portfolio demonstrations. Not intended for production-scale training.*\n\n[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/yourusername/llm-finetune-kit/blob/main/examples/01_quickstart.ipynb)\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fkaiser-data%2Fllm-finetune-kit","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fkaiser-data%2Fllm-finetune-kit","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fkaiser-data%2Fllm-finetune-kit/lists"}