More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    LitGPT: Train LLMs with Readable Code

    1

    Train and fine-tune LLMs with LitGPT's clean, hackable codebase: full fine-tuning, memory-efficient LoRA, pretraining, and deployment.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    LitGPT: Train LLMs with Readable Code

    LitGPT: Train LLMs with Readable Code

    Example session with this skill installed

    Fine-tune phi-2 on our support-ticket summaries with LoRA. I have one 16GB GPU and 3,000 examples.

    • Read your context and instructions
    • Compiled the litgpt

    Path chosen by situation: adapt a model to your data with limited VRAM — the LoRA row. Full fine-tuning wants 40GB+ for a 7B model; pretraining is for when no base model fits. LoRA on 16GB is the right call, and the smoke test came first: load and generate in Python before touching training, because if inference does not work, training will not either. Data prepared in Alpaca shape: instruction, input, output. The 3,000 examples were checked for the quality rules that matter more than quantity — every example looks like the real task, same input shape, same output style; template boilerplate, inconsistent formatting, and contradictory labels stripped out. A few thousand excellent examples beat fifty thousand mediocre ones. Run configured: litgpt finetune_lora microsoft/phi-2 --data JSON --data.json_path data/train.json --lora_r 16 --lora_alpha 32 --lora_dropout 0.05 --train.epochs 3 --train.learning_rate 1e-4 --train.micro_batch_size 4 --train.global_batch_size 32. Rank 16 as the default for a real adaptation task; query + value projections only, MLP layers left off without evidence. Watched early: loss curve in the first hundred steps — flat would mean the LR is too low for LoRA, exploding would mean it is too high. Loss fell cleanly. Deployment: adapters merged with litgpt merge_lora, then quantized to 8-bit after measuring the task metric before and after — the quality loss was small and measured, not assumed. The merged model serves like any HuggingFace checkpoint.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    LitGPT's pitch is simple: every model is readable code, and every workflow is a command. Use it when you want to understand what you are training, not just launch a job. This skill covers choosing your training path by situation and GPU reality — full fine-tuning, LoRA, pretraining from scratch, and the merge-to-deploy path — plus data preparation in Alpaca shape, where a few thousand excellent examples beat fifty thousand mediocre ones. LoRA tuning is concrete: rank selection by task complexity, which modules to adapt, and the learning rate that LoRA actually wants. From adapters to deployment, it covers merging, quantization measured against your task metric, and serving — plus a troubleshooting table for the failures every first run hits.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for LLM Ops.

    Maximum-scale pretraining (70B+) belongs to Megatron-style frameworks, not LitGPT.
    Inference only is a serving engine's job, not a training framework's.
    If you need YAML-config-driven experiment management, Axolotl or TRL fit better — this skill says so rather than stretching LitGPT.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before buying a skill.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 3 days ago

    Software
    Python

    Permissions

    Terminal / Shell
    Listed3 days ago

    What's inside

    Frequently Asked Questions

    LoRA, almost always. Full fine-tuning is for when adapters demonstrably underperform. Most people want the LoRA row.

    A few thousand excellent ones beat fifty thousand mediocre ones. Match the real task's input shape and output style, and strip template boilerplate and contradictory labels.

    The learning rate is too low for LoRA — raise it toward 1e-4. LoRA wants a higher LR than full fine-tuning; too low and the adapters never move.

    New in LLM Ops