Runnable recipes for fine-tuning, training, and deploying models on VESSL Cloud.
Each top-level folder is a self-contained recipe. Clone the repo (or just the folder you need), follow the recipe's README, and you'll have a working end-to-end run on VESSL Cloud.
| Recipe | Task | GPU | Approx. cost | Approx. time |
|---|---|---|---|---|
| gemma4-finetuning (reference implementation) | LoRA fine-tune Gemma 4 E4B on a small domain QA dataset | A100 SXM 80 GB × 1 | ~$0.43 | ~16 min |
| autoresearch | Run karpathy/autoresearch on cloud GPUs — an AI agent runs its own LLM pretraining experiments overnight, fanning out K candidates per round in parallel | H100 SXM 80 GB × 1 | $0.33/experiment ($5 / 16-experiment cycle) | ~8 min/experiment (~40 min / 16-experiment cycle) |
Prices as of 2026-05-03; see each recipe's benchmarks.md for details.
- Sign up for VESSL Cloud if you don't have an account.
- Pick a recipe folder and open its
README.md. - Follow either Path A (notebook in a workspace) or Path B (vesslctl batch job).
notebook/<recipe>.ipynb— interactive workspace walk-through. Good for a first run.batch-job/<recipe>.py+submit.sh— reproducible vesslctl invocation. Good for automation.data/+DATASET_CARD.md— bundled dataset with provenance and license.benchmarks.md— measured time, VRAM, cost, loss on VESSL Cloud.
- A VESSL Cloud account with credits.
- vesslctl installed and authenticated (required for Path B).
- A Hugging Face access token for gated models (Gemma 4 is gated).
- Open an issue describing the recipe (task, model, target GPU, expected cost/time).
- Copy the skeleton:
cp -r _template my-new-recipeand fill in the TODOs. The reference implementation isgemma4-finetuning/— match its section structure where it makes sense. - Run end-to-end on VESSL Cloud and record measured numbers in
benchmarks.md. - Open a PR.
AI coding assistants (Claude Code, Cursor, Codex, Aider) work well here — point them at the reference recipe + your _template/ copy. See CONTRIBUTING.md for a worked prompt and the full checklist.
- Code: Apache-2.0 (see LICENSE).
- Datasets: CC-BY-4.0 (declared per recipe in
data/DATASET_CARD.md).