sylvester-francis/typescript-slm-1.5b

TypeScript SLM 1.5B

A specialized Small Language Model for TypeScript code generation and understanding, optimized for React, Next.js, Angular, and Node.js frameworks.

Model Description

This model is a fine-tuned version of Qwen/Qwen2.5-Coder-1.5B-Instruct using LoRA (Low-Rank Adaptation) for parameter-efficient fine-tuning. It has been trained on 2,000-8,000 high-quality TypeScript code samples focusing on modern web development frameworks.

Key Features:

• Specialized in TypeScript and popular frameworks (React, Next.js, Angular, Node.js)
• Quality-scored training dataset with proper type annotations
• Optimized for code completion, generation, and understanding tasks
• Efficient inference with LoRA adapters

Intended Use

Primary Use Cases

• TypeScript code completion and generation
• React component scaffolding
• Next.js API route and page generation
• Angular service and directive creation
• Node.js/Express backend code generation
• Type definition and interface creation

Out-of-Scope Use

• Production-critical code generation without human review
• Non-TypeScript/JavaScript code generation
• General-purpose text generation
• Code obfuscation or malicious code generation

Training Data

The model was trained on a curated dataset of TypeScript code samples with the following distribution:

• React (43-58%): Components, hooks, context, custom hooks
• Angular (33-50%): Services, directives, modules, dependency injection
• Next.js (21-23%): Pages, API routes, SSR, SSG patterns
• TypeScript (9-16%): Advanced types, generics, utility types
• Node.js (6-11%): Express, NestJS, API servers

Dataset Quality Scoring:

• Samples scored 41-64 on quality metrics
• Prioritizes proper type annotations
• Excludes test files, debug code, and incomplete modules
• Focuses on production-quality patterns from popular repositories

Training Procedure

Training Hyperparameters

Hardware:

• Google Colab A100 40GB GPU
• CUDA acceleration with FP16 precision

Configuration:

• Base Model: Qwen/Qwen2.5-Coder-1.5B-Instruct
• Training Samples: 2,000-8,000 (depending on dataset tier)
• Epochs: 3
• Batch Size: 4
• Gradient Accumulation Steps: 8
• Effective Batch Size: 32
• Learning Rate: 2e-4
• Max Sequence Length: 1024
• LoRA Rank (r): 32
• LoRA Alpha: 16
• LoRA Dropout: 0.1
• Target Modules: All linear layers

Training Time:

• train_small.jsonl (2k samples): ~20-30 minutes on A100
• train_medium.jsonl (5k samples): ~50-75 minutes on A100
• train.jsonl (8k samples): ~2-3 hours on A100

Usage

Basic Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# Load base model and tokenizer
base_model = "Qwen/Qwen2.5-Coder-1.5B-Instruct"
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    device_map="auto",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained(base_model)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, "sylvester-francis/typescript-slm-1.5b")

# Generate code
prompt = """Write a React component that fetches user data and displays it in a card:

```typescript
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    do_sample=True,
    top_p=0.95
)

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

React Component Generation

prompt = """Create a TypeScript React component with props for a user profile card:

```typescript
interface UserProfileProps {
"""

# Generate with the model...

Next.js API Route

prompt = """Write a Next.js API route for user authentication:

```typescript
// pages/api/auth/login.ts
"""

# Generate with the model...

Angular Service

prompt = """Create an Angular service for HTTP data fetching:

```typescript
import { Injectable } from '@angular/core';
"""

# Generate with the model...

Performance

Code Quality Metrics

• Proper TypeScript type annotations
• Framework-specific best practices
• Adherence to modern ES6+ patterns
• Clean, readable code structure

Generation Speed

• Average: ~50-100 tokens/second on A100
• Latency: <100ms for typical completions
• Memory: ~3GB VRAM with adapter loaded

Limitations

1. Specialized Domain: Works best for TypeScript and related frameworks. Performance degrades for other languages.

2. Training Data Bias: Reflects patterns from popular open-source repositories, which may not match all coding styles.

3. Context Length: Limited to 1024 tokens, which may be insufficient for very large files.

4. No Real-time Updates: Training data is static and doesn't include the latest framework versions or patterns.

5. Requires Human Review: Generated code should always be reviewed for security, correctness, and best practices.

6. Type Safety: While trained on typed code, generated types may not always be complete or optimal.

Ethical Considerations

• Code Licensing: Ensure generated code complies with your project's license requirements
• Security: Always review generated code for security vulnerabilities
• Testing: Generated code should be thoroughly tested before production use
• Attribution: Consider the training data sources when using generated code commercially

Training Infrastructure

Software Stack:

• PyTorch 2.9.0+cu126
• Transformers 4.57.2
• PEFT 0.18.0
• TRL 0.25.1
• Datasets 4.0.0
• bitsandbytes 0.41.0+

Platform:

• Google Colab Pro (recommended)
• Supports Mac M4 (MPS) for local training (slower)
• Compatible with T4, A100, and other CUDA GPUs

Repository

Full training code, dataset filtering, and usage examples: https://github.com/sylvester-francis/slm-typescript-model

Model Card Authors

• Sylvester Francis (@sylvester-francis)

Citations

Base Model

@article{qwen2.5,
  title={Qwen2.5-Coder Technical Report},
  author={Qwen Team},
  year={2024},
  journal={arXiv preprint},
  url={https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct}
}

Training Framework

@misc{vonwerra2022trl,
  title={{TRL: Transformer Reinforcement Learning}},
  author={Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
  year={2020},
  journal={GitHub repository},
  publisher={GitHub},
  howpublished={\url{https://github.com/huggingface/trl}}
}

LoRA

@article{hu2021lora,
  title={LoRA: Low-Rank Adaptation of Large Language Models},
  author={Hu, Edward J and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu},
  journal={arXiv preprint arXiv:2106.09685},
  year={2021}
}

License

MIT License - See repository for full license text.

Acknowledgments

• Built on Qwen 2.5 Coder by Alibaba Cloud
• Training powered by Hugging Face TRL and PEFT libraries
• Dataset curated from high-quality open-source TypeScript projects