Yachay

Docs

Downloading your adapter

When your tune finishes, Yachay stages the weights and metadata to GCS and the dashboard shows three Download buttons — one per file. Each click mints a fresh 1-hour signed URL.

What you get

Three files, each as a separate signed download — no tarball:

  • adapter_model.safetensors

    The LoRA / QLoRA adapter weights. Loadable into HF transformers (PEFT), vLLM, or llama.cpp.

  • README.md

    Your declared model name, the base model + revision used, hyperparameters, and the YOU-OWN-THIS line.

  • LICENSE.txt

    License notice for the base model — Llama Community / Apache 2.0 / MIT / Gemma TOU depending on the base. Keep this with the adapter when you redistribute.

Loading it

Most teams use HuggingFace PEFT. The tokenizer comes from the base model — the adapter delivery does not include tokenizer files (the README records the exact base + revision used):

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "meta-llama/Llama-3.1-8B"  # see README.md for the exact base + revision

base = AutoModelForCausalLM.from_pretrained(BASE)
tokenizer = AutoTokenizer.from_pretrained(BASE)
model = PeftModel.from_pretrained(base, "./my-yachay-adapter")

# Optional: merge for single-file deployment
model = model.merge_and_unload()
model.save_pretrained("./my-merged-model")

vLLM and llama.cpp also support LoRA adapters directly — see each tool’s docs for the exact load syntax.

Retention lifecycle

  1. Day 0Tune completes. Adapter is staged to a GCS path scoped to your job ID. The dashboard shows three Download buttons (adapter, README, LICENSE).
  2. Days 1–30Free download access. Each click on the dashboard mints a fresh 1-hour signed URL — short by design so a stale URL pasted into chat can't be replayed.
  3. Day 30Adapter + dataset bytes are eligible for deletion. Extend retention from the job page at $0.06/GB-month + $0.50 handling fee per extension, up to 24 months. The Extend button shows the exact charge for the duration you pick before you commit — do this before the daily sweep runs.
  4. After Day 30A daily sweep purges bytes from GCS. Once the sweep has run, the extend control is replaced by a re-tune prompt — extensions can't restore data that no longer exists. The job record itself stays on your dashboard for billing history.
  5. Account deletionCascades — all adapters, datasets, and invoices for the deleted account are purged permanently. No soft-delete window.

Signed URL security

Signed URLs are scoped to a single file and expire in 1 hour. Each click on the dashboard Download button mints a fresh URL — if you paste an old one into Slack or email, it’s likely expired by the time anyone clicks. The short TTL is deliberate: an old URL pasted into chat can’t be replayed weeks later. Treat the URL as a secret anyway.

← All docs · Privacy