Skill UI

This guide details advanced methods for merging multiple specialized LLMs (e.g., combining math, coding, and chat capabilities) without the need for expensive retraining. Learn various techniques such as Linear Merging, SLERP, Task Arithmetic, TIES-Merging, and DARE. Use mergekit to create high-performing, specialized model variants quickly, significantly improving performance and reducing computational costs.

Automate the Continuous Integration and Continuous Deployment (CI/CD) workflow for applications leveraging Together AI's OpenAI-compatible API. This skill sets up comprehensive GitHub Actions workflows to run unit tests (mocking completions and embeddings) on every Pull Request, and full integration tests (live API calls) upon merging to the main branch. It ensures robust validation of prompt formatting, model selection, and API connectivity for various LLMs.

Open Dynamic Workflows (ODW) is an advanced, open-source engine designed to manage complex coding tasks using multiple AI agents. It operates in three stages: planning the task into parallel subtasks, orchestrating various coding agents (like Codex or Antigravity) to work concurrently, and finally, performing an adversarial verification pass to challenge and validate the combined output before merging. It supports multiple model providers and is ideal for structured, high-reliability software development.

Open Dynamic Workflows (ODW) is an advanced, open-source engine designed to manage complex coding tasks using multiple AI agents. It operates in three stages: planning the task into parallel subtasks, orchestrating various coding agents (like Codex or Antigravity) to work concurrently, and finally, performing an adversarial verification pass to challenge and validate the combined output before merging. It supports multiple model providers and is ideal for structured, high-reliability software development.

This skill covers compressing large language models using knowledge distillation techniques. It enables transferring capabilities from teacher models like GPT-4 to smaller student models, reducing inference costs while retaining performance. Key methods include temperature scaling, soft targets, reverse KLD, and logit distillation. Ideal for deploying efficient models and creating specialized versions.

Master techniques to extend transformer model context windows beyond original limits using RoPE, YaRN, and ALiBi. Ideal for processing long documents (32k-128k+ tokens) and implementing efficient positional encodings for LLMs. Covers rotary embeddings, attention biases, and interpolation strategies.

Learn to merge multiple fine-tuned models using mergekit without retraining. Covers methods like SLERP, TIES-Merging, and Task Arithmetic to combine domain expertise (math, coding, chat). Improves performance and reduces costs by blending capabilities efficiently. Ideal for creating specialized AI models and rapid experimentation.

This skill covers techniques to reduce LLM size and accelerate inference using methods like Wanda and SparseGPT. It enables model compression without retraining, achieving high sparsity with minimal accuracy loss. Suitable for deploying on constrained hardware like mobile or edge devices. Includes unstructured, structured, and N:M sparsity methods.

Learn to train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace. This skill covers sparse architectures like Mixtral 8x7B, reducing compute costs by 5x compared to dense models. It includes routing mechanisms, load balancing, expert parallelism, and inference optimization for scaling large-scale AI models efficiently.