LLM Training Frameworks
AI and machine learning skill, available on Zeplik
LLM Training Frameworks is a ready-to-run AI and machine learning skill on Zeplik. Training and fine-tuning large models — distributed pretraining (Megatron, FSDP, torchtitan), RLHF and MoE training, HF Transformers/Axolotl/NeMo, tokenizers and data pipelines. Ask in plain language and Zeplik applies the skill's method for you inside the conversation, on whichever AI model you prefer.
The LLM Training Frameworks skill loads automatically when your request matches it, or you can invoke it directly by typing /llm-training-frameworks in any chat. It works with attachments, connectors, and any model that supports the task, so you get the same expert method every time without setting anything up.
What the LLM Training Frameworks skill can do
- Recommend the right training method and framework for the goal
- Produce runnable configs and training code for LoRA, full, or RLHF runs
- Cover distributed pretraining setups like Megatron, FSDP, and torchtitan
- Guide MoE, RLHF, and tokenizer or data pipeline setup with rationale
Try these prompts on Zeplik
Pick a prompt to open it in the Zeplik app. If you are not signed in yet, your prompt is waiting for you the moment you do.
How the LLM Training Frameworks skill works
/llm-training-frameworks
Umbrella for training and fine-tuning large models. The user wants to train, fine-tune, or scale a model; establish the method (LoRA vs full vs pretraining vs RLHF), the framework, and the hardware, then deliver configs and training code plus a rationale. For inference route to llm-serving; for compression route to llm-quantization.
Dispatch table
Pick the reference file(s) that match the request, read them, then answer. Read at most 2-3 files per turn.
| Topic | Read |
|---|---|
| Expert guidance for fine-tuning LLMs with Axolotl - YAML configs, 100+ models, LoRA/QLo… | references/axolotl.md |
| Expert guidance for distributed training with DeepSpeed - ZeRO optimization stages, pip… | references/deepspeed.md |
| Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelis… | references/distributed-llm-pretraining-torchtitan.md |
| Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO… | references/fine-tuning-with-trl.md |
| Expert guidance for GRPO/RL fine-tuning with TRL for reasoning and task-specific model… | references/grpo-rl-training.md |
| Simplest distributed training API. | references/huggingface-accelerate.md |
| Fast tokenizers optimized for research and production. | references/huggingface-tokenizers.md |
| Implements and trains LLMs using Lightning AI's LitGPT with 20+ pretrained architecture… | references/implementing-llms-litgpt.md |
| Expert guidance for fine-tuning LLMs with LLaMA-Factory - WebUI no-code, 100+ models, 2… | references/llama-factory.md |
| Provides guidance for enterprise-grade RL training using miles, a production-ready fork… | references/miles-rl-training.md |
| Train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace. | references/moe-training.md |
| GPU-accelerated data curation for LLM training. | references/nemo-curator.md |
| High-performance RLHF framework with Ray+vLLM acceleration. | references/openrlhf-training.md |
| Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods. | references/peft-fine-tuning.md |
| Expert guidance for Fully Sharded Data Parallel training with PyTorch FSDP - parameter… | references/pytorch-fsdp.md |
| High-level PyTorch framework with Trainer class, automatic distributed training (DDP/FS… | references/pytorch-lightning.md |
| Scalable data processing for ML workloads. | references/ray-data.md |
| Distributed training orchestration across clusters. | references/ray-train.md |
| Language-independent tokenizer treating text as raw Unicode. | references/sentencepiece.md |
| Simple Preference Optimization for LLM alignment. | references/simpo-training.md |
| Provides guidance for LLM post-training with RL using slime, a Megatron+SGLang framework. | references/slime-rl-training.md |
| Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separa… | references/torchforge-rl-training.md |
| Trains large language models (2B-462B parameters) using NVIDIA Megatron-Core with advan… | references/training-llms-megatron.md |
| Expert guidance for fast fine-tuning with Unsloth - 2-5x faster training, 50-80% less m… | references/unsloth.md |
| Provides guidance for training LLMs with reinforcement learning using verl (Volcano Eng… | references/verl-rl-training.md |
| Deep learning framework (PyTorch Lightning). | references/pytorch-lightning-x.md |
How to work
- Identify which leaf topic the request maps to from the dispatch table above; establish the concrete inputs (language, dataset, framework, file format) and the goal. Ask for a missing detail rather than guessing.
- Read the matching reference file(s) before answering. Read at most 2-3 per turn.
- Deliver runnable artifacts — code, configs, specs — with a short rationale, matching the user's existing conventions when they paste code.
- Confirm any decision the source flags (versions, thresholds, tradeoffs) with the user instead of guessing.
Usage
/llm-training-frameworks $ARGUMENTS
How to use the LLM Training Frameworks skill
Sign in to Zeplik
Create a free Zeplik account or sign in. New accounts start with free credits, so you can try the LLM Training Frameworks skill right away.
Describe your AI and machine learning task
Ask in plain language, or type /llm-training-frameworks to invoke the skill directly. Zeplik recognizes the LLM Training Frameworks skill and applies its method.
Review and refine the result
Zeplik returns a clear, structured answer. Ask follow-ups in the same chat to refine it or take the next step.
Source and credit
- Author
- davila7 (D7 umbrella-consolidation)
- License
- MIT
Adapted from the open-source davila7/claude-code-templates project and tuned to run natively on Zeplik. View source on GitHub.
Frequently asked questions
- What is the LLM Training Frameworks skill?
- LLM Training Frameworks is a ready-to-run AI and machine learning skill on Zeplik. Training and fine-tuning large models — distributed pretraining (Megatron, FSDP, torchtitan), RLHF and MoE training, HF Transformers/Axolotl/NeMo, tokenizers and data pipelines. Ask in plain language and Zeplik applies the skill's method for you inside the conversation, on whichever AI model you prefer.
- How do I use LLM Training Frameworks on Zeplik?
- Sign in to Zeplik and ask in plain language, or type /llm-training-frameworks in any chat to invoke it directly. The skill applies its method and returns a result you can refine in the same conversation.
- Which AI model does the LLM Training Frameworks skill use?
- Any model you choose. Zeplik works across every model in one chat, so the LLM Training Frameworks skill runs on your preferred model for the task.
- Where does the LLM Training Frameworks skill come from?
- The LLM Training Frameworks skill is adapted from the open-source davila7/claude-code-templates project (MIT) and tuned to run natively on Zeplik. The original source is linked on this page.
- How much does the LLM Training Frameworks skill cost?
- Using the skill is free to start. You only spend Zeplik credits when the assistant runs, and new accounts begin with free credits.
Related ai and machine learning skills
- Agent MemoryMemory and context management for LLM agents — short/long-term memory stores, conversation persistence, context-window strategies (summarization, trimming, retrieval) and prompt caching. Use for "give my agent memory / manage context"; for vector stores see vector-databases.
- AI Safety GuardrailsSafety and moderation for LLM apps — Constitutional AI, Llama Guard input/output moderation, and NeMo Guardrails runtime rails. Use for "add safety/moderation/guardrails to my LLM app"; for evaluating safety see llm-evaluation-harnesses.
- Autonomous Agent PatternsFramework-agnostic design patterns for autonomous coding agents — goal decomposition, planning, parallel/multi-agent orchestration and operational modes. Use for "how should I architect an autonomous agent"; for concrete frameworks see ai-agent-frameworks.
- Gemini CLIRun Gemini CLI (Gemini 3 Pro) for code/plan review and huge >200k-token context analysis in the terminal
- InterpretabilityMechanistic interpretability of neural networks — TransformerLens, NNsight remote access, sparse autoencoders (SAELens), and causal interventions (pyvene). Use for "probe/intervene on model internals" or "train an SAE"; for architecture basics see llm-architectures.
- LLM App PatternsProduction patterns for LLM applications — RAG architecture, embeddings, LLMOps, and end-to-end app design. Use for "architect a production LLM/RAG app"; for the vector store see vector-databases, for agents see ai-agent-frameworks.
More on Zeplik
Try LLM Training Frameworks on Zeplik
Every model, one chat. Bring the LLM Training Frameworks skill into your next conversation and let the assistant do the work.