typescript-slm-1.5b-full/README.md

---
language:
- en
license: mit
library_name: transformers
tags:
- code
- typescript
- react
- nextjs
- angular
- nodejs
- qwen
- gguf
- ollama
base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
datasets:
- github-code
model-index:
- name: TypeScript-SLM-1.5B-Full
  results: []
---

# TypeScript-SLM-1.5B-Full

**TypeScript-SLM-1.5B** is a compact, domain-specialized language model fine-tuned for TypeScript code generation, with a focus on React, Next.js, Angular, and Node.js frameworks.

This repository contains the **full merged model** (base model + LoRA adapters) along with **GGUF quantized versions** ready for Ollama deployment.

## Model Description

- **Base Model**: [Qwen/Qwen2.5-Coder-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct)
- **Model Type**: Causal Language Model (Code Generation)
- **Parameters**: 1.5 billion
- **Context Length**: 1024 tokens
- **Fine-tuning Method**: LoRA (Low-Rank Adaptation)
- **License**: MIT
- **Language**: English (Code: TypeScript/JavaScript)

### Key Features

- ✅ Specialized in TypeScript code generation
- ✅ Framework-aware (React, Next.js, Angular, Node.js)
- ✅ Strongly-typed code with proper interfaces and types
- ✅ Optimized for modern web development patterns
- ✅ Available in multiple GGUF quantizations for Ollama
- ✅ Fast inference on consumer hardware

## Intended Uses

### Primary Use Cases

- **TypeScript Code Completion**: Auto-complete TypeScript code in IDEs
- **Component Generation**: Create React/Angular components from descriptions
- **Type Definition**: Generate TypeScript interfaces and type aliases
- **Code Snippets**: Quick generation of framework-specific patterns
- **Learning Aid**: Study TypeScript and framework best practices

### Example Prompts

```typescript
// React component with TypeScript
"Create a React component with TypeScript for a user profile card"

// Next.js API route
"Write a Next.js API route that handles user authentication"

// Angular service
"Create an Angular service for managing todo items"

// Node.js Express server
"Write an Express server with TypeScript that includes CORS and error handling"

// Type definitions
"Define a TypeScript interface for a blog post with author information"
```

## How to Use

### Option 1: Ollama (Recommended for Local Use)

The easiest way to use this model locally is with Ollama:

```bash
# Import the model using the Modelfile
ollama create typescript-slm-1.5b -f Modelfile-q4_k_m

# Run the model
ollama run typescript-slm-1.5b "Create a React component for a todo list"
```

**Available Quantizations:**
- `Modelfile-q4_k_m` - 4-bit quantization (~800MB, fastest)
- `Modelfile-q6_k` - 6-bit quantization (~1.2GB, balanced)
- `Modelfile-f16` - 16-bit float (~3GB, highest quality)

### Option 2: Transformers (Python)

Use directly with the Transformers library:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "sylvester-francis/typescript-slm-1.5b-full"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Generate code
prompt = "Create a React component with TypeScript for a user profile card:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.3,
    top_p=0.95,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)

code = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(code)
```

### Option 3: GGUF Files (llama.cpp)

Download GGUF files directly for use with llama.cpp:

```bash
# Download specific quantization
huggingface-cli download sylvester-francis/typescript-slm-1.5b-full \
  gguf/typescript-slm-1.5b-q4_k_m.gguf --local-dir ./models

# Run with llama.cpp
./llama-cli -m ./models/gguf/typescript-slm-1.5b-q4_k_m.gguf \
  -p "Create a TypeScript interface for a user profile"
```

## Model Details

### Architecture

- **Base**: Qwen2.5-Coder-1.5B-Instruct
- **Modifications**: LoRA fine-tuning on TypeScript-specific data
- **LoRA Rank**: 64
- **LoRA Alpha**: 128
- **Target Modules**: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

### Training Data

The model was fine-tuned on a curated dataset of TypeScript code:

- **Sources**: GitHub repositories (1000+ stars), StackOverflow Q&A
- **Total Samples**: ~8,000 high-quality code samples
- **Quality Filtering**: Intelligent scoring based on TypeScript features
- **Framework Distribution**:
  - React: ~50% (components, hooks, context)
  - Angular: ~25% (services, directives, modules)
  - Next.js: ~15% (pages, API routes, SSR)
  - Node.js: ~10% (Express, NestJS, APIs)

**Quality Indicators:**
- Proper TypeScript type annotations
- Complete modules with imports/exports
- Framework-specific best practices
- Production-quality code from popular repositories

### Training Configuration

```yaml
Base Model: Qwen/Qwen2.5-Coder-1.5B-Instruct
Training Method: LoRA Fine-tuning
LoRA Rank: 64
LoRA Alpha: 128
Learning Rate: 2e-4
Batch Size: 8 (effective: 32 with gradient accumulation)
Epochs: 3
Max Sequence Length: 1024
Optimizer: AdamW
Warmup Ratio: 0.03
```

### Training Hardware

- **Platform**: Google Colab A100 (40GB)
- **Training Time**: ~30 minutes
- **Framework**: Hugging Face TRL + PEFT

## Performance

### Code Quality Metrics

Based on evaluation of 100 TypeScript generation tasks:

| Metric | Score |
|--------|-------|
| **Correct Syntax** | 85% |
| **Proper TypeScript Types** | 72% |
| **Framework Best Practices** | 68% |
| **Context Understanding** | 1024 tokens |

### Generation Speed

| Platform | Tokens/Second |
|----------|---------------|
| NVIDIA RTX 3090 (FP16) | ~80 tokens/s |
| Apple M2 Max (GGUF q4_k_m) | ~45 tokens/s |
| Apple M1 (GGUF q4_k_m) | ~30 tokens/s |
| CPU (GGUF q4_k_m) | ~10-15 tokens/s |

## Repository Contents

```
typescript-slm-1.5b-full/
├── config.json                           # Model configuration
├── tokenizer.json                        # Tokenizer configuration
├── tokenizer_config.json                 # Tokenizer settings
├── generation_config.json                # Generation parameters
├── pytorch_model.bin                     # Full merged PyTorch model
├── model.safetensors                     # SafeTensors format
│
└── gguf/                                 # GGUF quantized models
    ├── typescript-slm-1.5b-q4_k_m.gguf  # 4-bit quantization (~800MB)
    ├── typescript-slm-1.5b-q6_k.gguf    # 6-bit quantization (~1.2GB)
    ├── typescript-slm-1.5b-f16.gguf     # 16-bit float (~3GB)
    ├── Modelfile-q4_k_m                  # Ollama config (4-bit)
    ├── Modelfile-q6_k                    # Ollama config (6-bit)
    └── Modelfile-f16                     # Ollama config (16-bit)
```

## Quantization Comparison

| Format | Size | Quality | Speed | Use Case |
|--------|------|---------|-------|----------|
| **q4_k_m** | ~800MB | Good | Fastest | Local development, quick iteration |
| **q6_k** | ~1.2GB | Very Good | Fast | Production, balanced performance |
| **f16** | ~3GB | Excellent | Moderate | High quality, benchmarking |
| **PyTorch** | ~3GB | Perfect | GPU-dependent | Fine-tuning, research |

**Recommendation**: Use `q4_k_m` for testing, `q6_k` for production.

## Limitations

### Known Limitations

1. **Context Length**: Limited to 1024 tokens (larger contexts may lose coherence)
2. **Complex Logic**: May struggle with very complex algorithmic tasks
3. **Newer Frameworks**: Limited knowledge of frameworks released after training cutoff
4. **Type Inference**: Sometimes requires explicit type annotations
5. **Edge Cases**: May not handle all TypeScript edge cases correctly

### Not Recommended For

- ❌ Production-critical code without review
- ❌ Security-sensitive implementations
- ❌ Complex algorithm design
- ❌ Large-scale refactoring
- ❌ Framework versions beyond training data

### Best Used With

- ✅ Human review and validation
- ✅ Existing codebase context
- ✅ Clear, specific prompts
- ✅ Common framework patterns
- ✅ Learning and prototyping

## Ethical Considerations

### Intended Use

This model is designed for:
- **Developer productivity**: Assisting professional developers
- **Learning**: Helping students learn TypeScript and frameworks
- **Prototyping**: Quick generation of boilerplate code

### Potential Risks

- **Code Quality**: Generated code should always be reviewed
- **Security**: May generate insecure patterns if prompted
- **Licensing**: Generated code may resemble training data patterns
- **Bias**: May reflect patterns common in open-source code

### Responsible Use Guidelines

1. **Always Review**: Never deploy generated code without review
2. **Test Thoroughly**: All generated code should be tested
3. **Check Licenses**: Ensure compliance with relevant licenses
4. **Security Audit**: Review for security vulnerabilities
5. **Attribution**: Credit the model when appropriate

## Related Models

### Model Family

- **[typescript-slm-1.5b](https://huggingface.co/sylvester-francis/typescript-slm-1.5b)** - LoRA adapter only
- **[typescript-slm-7b](https://huggingface.co/sylvester-francis/typescript-slm-7b)** - Larger 7B variant
- **[typescript-slm-7b-reasoning](https://huggingface.co/sylvester-francis/typescript-slm-7b-reasoning)** - 7B with enhanced reasoning

### Comparison

| Model | Parameters | Context | Speed | Quality | Use Case |
|-------|------------|---------|-------|---------|----------|
| **1.5B** | 1.5B | 1024 | Fastest | Good | Local dev, quick iteration |
| **7B** | 7B | 2048 | Fast | Excellent | Production code |
| **7B-Reasoning** | 7B | 2048 | Moderate | Excellent+ | Complex problems, debugging |

## Training Pipeline

This model was created using the [TypeScript SLM training pipeline](https://github.com/sylvester-francis/slm-typescript-model):

```bash
# Train your own model
git clone https://github.com/sylvester-francis/slm-typescript-model
cd slm-typescript-model
pip install -r requirements.txt

# Run complete pipeline
python slm.py pipeline

# Deploy to Ollama
python slm.py deploy typescript-slm-1.5b
```

## Citation

If you use this model in your research or project, please cite:

```bibtex
@software{typescript_slm_1.5b_2025,
  author = {Francis, Sylvester},
  title = {TypeScript-SLM-1.5B: Domain-Specialized Language Model for TypeScript},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/sylvester-francis/typescript-slm-1.5b-full}
}
```

## Acknowledgments

- **Base Model**: [Qwen2.5-Coder-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct) by Alibaba Cloud
- **Training Framework**: [Hugging Face TRL](https://github.com/huggingface/trl) and [PEFT](https://github.com/huggingface/peft)
- **GGUF Conversion**: [llama.cpp](https://github.com/ggerganov/llama.cpp) by Georgi Gerganov

## License

This model is released under the **MIT License**.

```
MIT License

Copyright (c) 2025 Sylvester Francis

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
```

## Contact & Support

- **Issues**: https://github.com/sylvester-francis/slm-typescript-model/issues
- **Repository**: https://github.com/sylvester-francis/slm-typescript-model
- **HuggingFace**: https://huggingface.co/sylvester-francis

## Version History

- **v1.0.0** (2025-11-29): Initial release
  - Full merged model with LoRA adapters
  - GGUF quantizations (q4_k_m, q6_k, f16)
  - Ollama Modelfiles
  - Optimized for React, Next.js, Angular, Node.js

---

**Keywords**: typescript, react, nextjs, angular, nodejs, code-generation, llm, gguf, ollama, qwen, lora, small-language-model