More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    LLaMA-Factory: Fine-Tune LLMs No-Code or by YAML

    1

    Fine-tune LLMs through LLaMA-Factory's unified interface — WebUI for no-code runs, CLI/YAML for repeatable ones, plus export.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    LLaMA-Factory: Fine-Tune LLMs No-Code or by YAML

    LLaMA-Factory: Fine-Tune LLMs No-Code or by YAML

    Example session with this skill installed

    Fine-tune Llama-3-8B-Instruct on our support data with QLoRA. I want a repeatable run I can re-execute, not a one-off click session.

    • Read your context and instructions
    • Compiled the llama-factory

    Setup verified by CLI first: llamafactory-cli train --help printed cleanly before the WebUI was ever opened. GPU floor checked — 16GB+ VRAM for 7B-class QLoRA; CPU-only is not a training target, and the run was not attempted on one. Prototype in the WebUI, promote to YAML: the first exploratory run was clicked through in the WebUI, and the moment it looked promising its CLI command was saved — then the real run lived in a version-controlled YAML file. A run that cannot be re-executed from a file did not happen. Data: support records formatted as alpaca in train.json, described in data/dataset_info.json with instruction/input/response columns. 1k–10k high-quality examples, matching the deployment input shape exactly (system+user turns in production means system+user turns in training), with a 5–10% validation slice held out. Configuration: model meta-llama/Meta-Llama-3-8B-Instruct, method lora, 4-bit quantization, lora_rank 16, alpha 32, dropout 0.05, targets q_proj and v_proj, learning rate 5.0e-5, 3 epochs, per-device batch 2, accumulation 8. Watched three things: train and eval loss (eval rising while train falls would mean overfitting), gradient norms (spikes mean instability), and sample outputs mid-training on held-out prompts — outputs lie least of all. Evaluation: task metric on the held-out set before and after, not loss alone. Export: merged into the base model with llamafactory-cli export, then verified the merged model generates sane output before deploying — a bad merge is silent.

    support records formatted as alpaca in train.json

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

    About this skill

    LLaMA-Factory wraps the standard fine-tuning stack in one consistent interface: point it at a model, describe your data, pick a method, train. It supports 100+ models, LoRA and QLoRA across bit-widths, and vision-language tuning. The skill teaches the two ways to drive it and the rule that connects them: prototype in the WebUI, promote to YAML, because a run you cannot re-execute from a file did not happen. Data formatting is concrete — JSON datasets described in dataset_info.json, alpaca or sharegpt formatting, a held-out validation slice — and the QLoRA configuration is spelled out: bit-width choice, rank, target modules, learning rate, batching. Training is watched on three signals — loss curves, gradient norms, and sample outputs — and evaluation means a task metric before and after, not 'loss went down.' Export covers adapters or merged weights, with the warning that a bad merge is silent.

    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.

    GPU with 16GB+ VRAM is the practical floor for 7B-class QLoRA; CPU-only is not a training target.
    Training loss alone lies — the validation slice and the task metric are the verdict, not the curve.
    A bad merge is silent: the merged model must be verified to generate sane output before deployment.

    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
    Git

    Permissions

    Terminal / Shell
    Listed3 days ago

    What's inside

    Frequently Asked Questions

    Prototype in the WebUI, promote to YAML. The WebUI is good for first runs and exploring hyperparameters; serious runs live in version-controlled configs you can re-execute.

    1k–10k high-quality examples beat 100k noisy ones. Match the deployment input shape exactly and hold out 5–10% for validation.

    Three things: loss curves (train and eval), gradient norms, and sample outputs mid-training on held-out prompts. Numbers lie less often than loss curves, but outputs lie least of all.

    New in LLM Ops