Expert TypeScript

TypeScript code generation and instruction expert trained on multiple high-quality sources including synthetic datasets, official TypeScript documentation, and real-world code samples.

Quick Start

# 1. Download package (when available)
# wget https://github.com/hivellm/expert-typescript/releases/download/v0.0.1/expert-typescript-qwen3-0-6b.v0.0.1.expert

# 2. Install
expert-cli install expert-typescript-qwen3-0-6b.v0.0.1.expert

# 3. Use
expert-cli chat --experts typescript
> Create a function that validates email addresses using regex

Works best for: Code generation, refactoring, type definitions, modern TypeScript patterns
Limitations: See Known Limitations below

Features

• ✅ Code Generation: Generate TypeScript code from natural language descriptions
• ✅ Code Explanation: Explain TypeScript code logic and patterns
• ✅ Refactoring: Suggest improvements and refactorings for TypeScript code
• ✅ Type Safety: Understand and apply TypeScript type system
• ✅ Modern Patterns: Knowledge of modern TypeScript features and best practices
• ✅ DoRA adapter (r=12) optimized for code generation tasks
• ✅ 207,283 validated examples from official docs and open-source TypeScript code
• ✅ Growing dataset combining TypeScript Handbook extractions and large curated slices of the-stack TypeScript subset

What It Can Do ✅

Excellent Support:

• ✅ Function generation from natural language
• ✅ Interface and type definitions
• ✅ Class implementations
• ✅ Code refactoring suggestions
• ✅ Code explanation and documentation
• ✅ Modern TypeScript features (generics, async/await, decorators)
• ✅ Type inference and type safety
• ✅ Common patterns (singletons, factories, builders)

Example Outputs:

// Code Generation
Input: "Create a function that validates email addresses using regex"
Output: 
function validateEmail(email: string): boolean {
  const emailRegex = /^[^\s@]+@[^\s@]+\.[^\s@]+$/;
  return emailRegex.test(email);
}

// Type Generation
Input: "Generate TypeScript interface for a user profile"
Output:
interface UserProfile {
  id: string;
  username: string;
  email: string;
  createdAt: Date;
  isActive: boolean;
  roles: string[];
}

// Code Explanation
Input: "Explain the logic of the useDispatcher function"
Output: The useDispatcher function takes an argument that can be a Dispatcher, 
DispatcherRegisterOptions, or an array of these types. It registers dispatchers 
with the store system...

Known Limitations ⚠️

These patterns may have lower success rates:

• ⚠️ Very complex type inference - May struggle with deeply nested generics
• ⚠️ Advanced decorators - Limited support for complex decorator patterns
• ⚠️ Framework-specific code - May not follow specific framework conventions perfectly
• ⚠️ Large codebases - Best for single functions/interfaces, not entire modules

Recommendation:

• ✅ Use for generating individual functions, types, and small code snippets
• ✅ Provide clear, specific instructions for best results
• ⚠️ Review generated code for framework-specific patterns

Dataset

Current Dataset

• Total Examples: 207,283 instruction-output pairs
• Sources:
  • TypeScript Handbook extraction (scripts/extract_typescript_docs.py): 155 examples
  • bigcode/the-stack TypeScript subset (scripts/integrate_the_stack.py --limit 100000): 207,128 examples
• Format: Instruction-tuning dataset for TypeScript code generation (ChatML format)
• Preprocessing: Deduplication applied (50,026 duplicates removed)
• Location: datasets/train.jsonl
• Integration Date: 2025-11-08

Dataset Statistics

• Expansion: Initial merge of documentation + curated The Stack subset (26k examples) expanded to ~207k examples
• Quality: Blend of official documentation snippets and a broad sample of real-world code patterns
• Coverage: Core TypeScript syntax plus significant representation of advanced language features and patterns

Training

Quick Start

# From expert-typescript directory
cd F:/Node/hivellm/expert/experts/expert-typescript

# Run training (CLI will download dataset automatically)
../../cli/target/release/expert-cli train

Training Configuration

• Base Model: Qwen3-0.6B (INT4 quantized)
• Adapter: DoRA (rank=12, alpha=24)
• Target Modules: q_proj, k_proj, v_proj, o_proj, up_proj, down_proj
• Epochs: 3
• Learning Rate: 0.0003
• Batch Size: 16
• Gradient Accumulation: 4 steps
• Max Sequence Length: 2048

Dataset Information

The dataset currently includes:

• TypeScript Handbook documentation examples (155)
• Curated the-stack TypeScript code samples (207,128)

All datasets have been merged into datasets/train.jsonl with deduplication applied. The dataset is ready for additional augmentation and training experiments.

Testing

Run the automated tests to validate the expert:

# Windows
.\test.ps1

# Linux/macOS
./test.sh

Qualitative Checkpoint Review

Generate side-by-side outputs for the base model and every available checkpoint:

cd F:/Node/hivellm/expert/experts/expert-typescript
F:/Node/hivellm/expert/cli/venv_windows/Scripts/python.exe compare.py

The script prints formatted outputs for each representative TypeScript task and saves a JSON summary as checkpoint_comparison_results.json for further analysis.

Usage Examples

Code Generation

Input: "Create a function that validates email addresses using regex"
Output: 
function validateEmail(email: string): boolean {
  const emailRegex = /^[^\s@]+@[^\s@]+\.[^\s@]+$/;
  return emailRegex.test(email);
}

Code Explanation

Input: "Explain the logic of the useDispatcher function"
Output: The useDispatcher function takes an argument that can be a Dispatcher, 
DispatcherRegisterOptions, or an array of these types. It registers dispatchers 
with the store system...

Type Generation

Input: "Generate TypeScript interface for a user profile"
Output:
interface UserProfile {
  id: string;
  username: string;
  email: string;
  createdAt: Date;
  isActive: boolean;
  roles: string[];
}

Python Integration

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model_path = "F:/Node/hivellm/expert/models/Qwen3-0.6B"
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_path,
    device_map="auto",
    dtype=torch.bfloat16,
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained(base_model_path, trust_remote_code=True)

# Load expert adapter
adapter_path = "experts/expert-typescript"
model = PeftModel.from_pretrained(base_model, adapter_path)

# Generate TypeScript code
prompt = "Create a function that validates email addresses using regex"

messages = [
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=200,
        temperature=0.7,
        top_p=0.8,
        top_k=20
    )

code = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(code)

Package

Once trained, package the expert:

expert-cli package \
  --manifest manifest.json \
  --weights weights/qwen3-06b/adapter \
  --output expert-typescript-qwen3-0-6b.v0.0.1.expert

Performance Metrics

Metric	Current
Adapter Size	~26 MB
VRAM Overhead	~18 MB
Load Time	<10ms (hot)
Training Time	~8-12 hours (RTX 4090)
Dataset Size	207,283 examples

License

MIT License - Same as the source dataset

Credits

• Datasets:
  • mhhmm/typescript-instruct-20k-v2c (19,994 examples)
  • TypeScript Handbook (Microsoft) - Official documentation (155 examples)
  • bigcode/the-stack - TypeScript subset (298,543 examples)
• Base Model: Qwen3-0.6B by Alibaba Cloud
• Training Framework: HuggingFace PEFT + Transformers