SLMGEN is a web application that automates SLM fine-tuning. Upload your JSONL dataset and receive ready-to-run Google Colab notebooks with Unsloth + LoRA optimization.
Your Data → Best Model → Matched. One notebook. Zero setup. Ready to train.
| Feature | Description |
|---|---|
| 📤 Smart Upload | Drag-and-drop JSONL with Live Chat Preview (min 50 examples) |
| 📊 Quality Scoring | Duplicate detection, consistency checks, 0-100% quality score |
| 🧠 18 Model Support | Qwen 3.5, Llama 3.3, DeepSeek V3, Phi-4, Gemma 3, SmolLM3 + more |
| 🎯 100-Point Matching | Task fit (50pts) + Deploy target (30pts) + Data traits (20pts) |
| 💻 Training Simulator | Real-time terminal simulation during generation phase |
| 📓 Self-Contained Notebooks | Dataset embedded as base64 - no file uploads needed |
| 🔄 Dataset Converter | CSV, TSV, JSON, Alpaca, ShareGPT → ChatML |
| ⚡ Training Presets | Quick Demo, Production, Edge, Code, Long Context |
| 📦 Export Options | Ollama, GGUF, vLLM, HuggingFace |
- Personality Detection - Infers tone, verbosity, technicality, strictness
- Hallucination Risk - Scores likelihood of model fabrication (0-1)
- Confidence Score - Measures training reliability via coverage/diversity
- Behavior Composer - Generate system prompts from trait sliders
- Prompt Linter - Detects contradictions, redundancy, ambiguity
- Prompt Diff - Semantic comparison between prompts
- "Why This Model?" - Strength/weakness deep dive per model
- Failure Previews - Synthetic failure cases before training
- Model Card Generator - Auto-generated deployment README
| Component | Technology |
|---|---|
| Backend | Python 3.11, FastAPI, Pydantic v2 |
| Session Store | In-Memory (Thread-safe TTL eviction, zero Redis dependency) |
| Frontend | Next.js 16, TypeScript, React 19, Framer Motion |
| Design | Tailwind CSS, JetBrains Mono, Everblush Theme |
| Auth | Supabase (Optional OAuth + Email, or local mock) |
| Training | Unsloth + LoRA on Google Colab (Free T4 & A100 tiers) |
| Deployment | Vercel (Frontend) + Render (Backend) |
- Python 3.11+ (or uv)
- Node.js 18+
- Supabase project (optional — set
AUTH_DISABLED=truefor 100% local development without Supabase)
cd libslmgen
# Option A: Instant run with uv (recommended)
uv run uvicorn app.main:app --reload --port 8000
# Option B: Standard virtualenv
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
uvicorn app.main:app --reload --port 8000cd slmgenui
npm install
cp .env.example .env.local # Configure API URL + Supabase
npm run devOpen http://localhost:3000 🎉
slmgen/
├── libslmgen/ # Python Backend
│ ├── app/
│ │ ├── main.py # FastAPI app
│ │ ├── session_store.py # Thread-safe in-memory session store
│ │ ├── models.py # Pydantic data schemas
│ │ ├── config.py # Environment & settings
│ │ └── routers/ # API endpoints
│ │ ├── upload.py # Dataset upload & validation
│ │ ├── analyze.py # Dataset analysis
│ │ ├── recommend.py # Model recommendation
│ │ ├── generate.py # Notebook generation
│ │ ├── convert.py # Dataset format conversion (CSV, JSON, Alpaca, ShareGPT)
│ │ ├── presets.py # Training presets (Quick Demo, Production, etc.)
│ │ ├── export.py # Model export guides (Ollama, GGUF, vLLM)
│ │ ├── advanced.py # Intelligence features
│ │ └── jobs.py # Job history (Supabase)
│ └── core/
│ ├── ingest.py # JSONL parsing & validation
│ ├── quality.py # Quality scoring
│ ├── analyzer.py # Dataset analysis
│ ├── recommender.py # 100-point scoring engine + GPU tiering
│ ├── notebook.py # Jupyter notebook generator
│ ├── convert.py # Format converters
│ ├── training_presets.py # Hyperparameter presets
│ ├── export.py # Export template generator
│ ├── personality.py # Personality detection
│ ├── risk.py # Hallucination risk
│ ├── confidence.py # Training confidence
│ ├── behavior.py # Behavior composer
│ ├── prompt_linter.py # Prompt linting
│ └── model_card.py # README generator
├── slmgenui/ # Next.js 16 Frontend
│ └── src/
│ ├── app/ # Pages (dashboard, login, signup, history, settings)
│ ├── components/ # UI components & charts
│ ├── lib/ # API client & types
│ └── hooks/ # React hooks (with sessionStorage persistence)
├── docs/
│ ├── API.md # API reference
│ ├── USER_GUIDE.md # User guide
│ └── DEPLOY.md # Deployment guide
└── supabase/
└── schema.sql # Database schema
| Model | Size | Context | Colab GPU Tier | Best For | Gated |
|---|---|---|---|---|---|
| DeepSeek V3 | 84B | 64K | A100 (Pro) | MoE reasoning, complex QA | ❌ |
| Llama 3.3 70B | 70B | 128K | A100 (Pro) | SOTA quality, reasoning | ✅ |
| Qwen 3.5 32B | 32B | 64K | A100 (Pro) | Hybrid thinking, coding | ❌ |
| Mistral Small 3 | 24B | 131K | A100 (Pro) | Code, 128K context | ❌ |
| Qwen 2.5 14B | 14B | 32K | A100 (Pro) | Long context, reasoning | ❌ |
| Llama 3.3 8B | 8B | 128K | T4 (Free) | General purpose, 128K context | ✅ |
| Mistral 7B | 7B | 32K | T4 (Free) | Creative generation, QA | ❌ |
| Qwen 3 4B | 4B | 32K | T4 (Free) | Thinking mode, math, code | ❌ |
| Gemma 3 4B | 4B | 131K | T4 (Free) | Multimodal, long context | ✅ |
| Phi-4 Mini | 3.8B | 16K | T4 (Free) | Classification, extraction | ❌ |
| SmolLM3 3B | 3B | 128K | T4 (Free) | Multilingual, edge-ready | ❌ |
| Llama 3.2 3B | 3B | 8K | T4 (Free) | Fast Q&A, conversations | ✅ |
| Qwen 2.5 3B | 3B | 32K | T4 (Free) | Multilingual, JSON output | ❌ |
| Gemma 2 2B | 2B | 8K | T4 (Free) | Edge, mobile, browser | ✅ |
| SmolLM2 1.7B | 1.7B | 8K | T4 (Free) | Ultra-compact, low memory | ❌ |
| Llama 3.2 1B | 1B | 8K | T4 (Free) | Lightweight mobile | ✅ |
| TinyLlama | 1.1B | 2K | T4 (Free) | Minimal compute demos | ❌ |
Each line in your JSONL file should be a conversation:
{"messages": [{"role": "user", "content": "Hello!"}, {"role": "assistant", "content": "Hi there!"}]}
{"messages": [{"role": "system", "content": "You are helpful."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}Requirements:
- ✅ Minimum 50 examples
- ✅ At least one user and one assistant message
- ✅ UTF-8 encoding
- ✅ Valid JSON per line
npx vercel --prodUses render.yaml blueprint for auto-deployment.
See DEPLOY.md for full instructions.
# Backend (.env)
ALLOWED_ORIGINS=https://slmgen.vercel.app,http://localhost:3000
SESSION_TTL_SECONDS=1800
AUTH_DISABLED=true # Set to true for zero-setup local dev without Supabase
# Optional: Supabase (for persistent job history and user authentication)
SUPABASE_URL=your_supabase_url
SUPABASE_ANON_KEY=your_anon_key
SUPABASE_SERVICE_KEY=your_service_key
SUPABASE_JWT_SECRET=your_jwt_secret
# Optional: HuggingFace Token (for validating gated models like Llama/Gemma)
HF_TOKEN=hf_...
# Frontend (.env.local)
NEXT_PUBLIC_API_URL=http://localhost:8000
NEXT_PUBLIC_SUPABASE_URL=your_supabase_url
NEXT_PUBLIC_SUPABASE_ANON_KEY=your_anon_keyMIT License - See LICENSE
Vedant Singh Rajput
Eshan Roy
⭐ Star this repo if SLMGEN helped you fine-tune faster!