More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Unsloth Speedrun: Fine-Tune Fast on Small GPUs

    1

    Fine-tune open language models fast and on small GPUs using Unsloth's optimized LoRA and QLoRA kernels — memory math, a minimal recipe, and the failure modes to watch.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Unsloth Speedrun: Fine-Tune Fast on Small GPUs

    Unsloth Speedrun: Fine-Tune Fast on Small GPUs

    Example session with this skill installed

    We have a 24GB GPU and a 7B base model. Fine-tune it on 2,000 support replies so it answers in our voice, without buying hardware.

    • Read your context and instructions
    • Compiled the unsloth speedrun

    Did the memory math first: a 7B model in 4-bit quantization fits the weights in about 4GB, the optimizer states stay small with an 8-bit optimizer, and a rank-16 adapter adds only tens of megabytes — the whole run fits on the 24GB card with room for a reasonable batch. Picked the base model for the job, not the leaderboard: an instruction-tuned 7B with a context window that fits the longest support reply, and a license that allows commercial use — checked before training, not after. Prepared the data as instruction pairs from the 2,000 replies, cleaned of personal information and deduped, with 100 held out for evaluation so the training set never grades itself. Ran the recipe: 4-bit base, rank-16 adapter, three epochs, and a learning rate the recipe recommends for this size — and watched the loss curve plus a real check: sample replies on the held-out set after each epoch. The loss fell smoothly, but the epoch-two samples were already repeating phrases from the training data, so training stopped at two epochs instead of three — the samples caught what the loss curve hid. The deliverable is the adapter, the training config, the held-out evaluation with before-and-after samples, and the plain statement of what the model does not do: it answers in the house voice on covered topics, and it still says it does not know on the rest.

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

    About this skill

    Fine-tuning a language model used to mean a large GPU and a long wait. This skill fine-tunes open language models fast and on small GPUs using Unsloth's optimized LoRA and QLoRA kernels — memory math, a minimal training recipe, quantization choices, and the failure modes to watch. It covers the recipe end to end: picking the base model, sizing the adapter, preparing the data, and running the training loop without running out of memory. It also covers the honest part: when fine-tuning is the wrong tool, how to tell the adapter actually learned anything, and what to do when the loss curve lies.

    It covers Unsloth's LoRA and QLoRA path; full fine-tuning and other frameworks need their own recipes.
    A fine-tune inherits the base model's limits — it will not fix a base model that cannot do the task at all.
    Loss curves can lie. The skill pairs every run with held-out samples, because a falling loss and a memorizing model look the same on a chart.

    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 any payment.

    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

    Permissions

    No special permissions declared or detected

    Listed3 days ago

    What's inside

    Frequently Asked Questions

    Yes, with 4-bit quantization for the weights, an 8-bit optimizer, and a LoRA adapter — the skill shows the memory math before you start, so there are no surprises mid-run.

    Watch held-out samples, not just the loss. When the samples start repeating training phrases, the model is memorizing — stop there, even if the loss is still falling.

    Only if the base model is close but the voice, format, or domain is off. If the base model cannot do the task, fine-tuning will not teach it — pick a better base or a different approach.

    New in LLM Ops