Addition of 5 Nosana Templates across different categories

[Nikhil4615]

PR: Addition of 5 Nosana Templates across different categories

Overview

1. DeepCoder-14B-Preview

State-of-the-art code generation model

• 60.6% Pass@1 on LiveCodeBench v5 (comparable to OpenAI's o3-mini)
• 64K context window for handling complex programming tasks
• 1936 Codeforces Rating (95.3 percentile) demonstrates superior code quality
• Efficient 14.8B parameter design with MIT License for commercial use

2. XLM-RoBERTa Large

Powerful multilingual transformer for cross-lingual understanding

• Pretrained on 2.5TB of data across 100 languages
• 561M parameters optimized for multilingual tasks
• Excellent zero-shot cross-lingual transfer capabilities
• Ideal foundation for fine-tuning classification and token tagging

3. Mixtral-8x7B-Instruct

Advanced Mixture-of-Experts language model with instruction tuning

• 46.7B total parameters with 12.9B active parameters for efficiency
• Outperforms Llama 2 70B on multiple benchmarks
• 32K context window for handling complex tasks
• Support for 5 languages with strong reasoning capabilities

4. MobileNetV3-Small

Ultra-efficient vision model for edge deployment

• Just 2.5M parameters with minimal computational requirements (0.1 GMACs)
• Optimized for mobile and IoT devices with small memory footprint
• Pre-trained on ImageNet for reliable image classification
• Deployable on devices with minimal GPU resources

5. Whisper Large V3

State-of-the-art speech recognition and transcription model

• Supports 99 languages with 10-20% error reduction from previous versions
• 1.5B parameter transformer-based architecture
• Advanced timestamp capabilities for precise audio alignment
• New Cantonese language support in this version

What this PR consists?

• Each template includes standardized files (info.json, job-definition.json, README.md)
• Templates use appropriate container technologies (vLLM, TorchServe, Text Embeddings Inference)
• All models expose RESTful APIs or OpenAI-compatible endpoints
• GPU memory requirements clearly documented for deployment planning

This diverse collection spans multiple AI domains - from speech recognition and multilingual text processing to efficient vision models, multimodal reasoning, and code generation - all unified through a standardized deployment framework.