Release notes¶
What changed in each version of Tuneline Studio, newest first.
Version 0.1.0 (September 2026)¶
The first public release of Tuneline Studio: build fine-tuning pipelines on a canvas, run them on free or paid compute with your own keys, and follow every run live.
Features¶
Templates. Start from a ready-made pipeline on the New pipeline screen:
| Template | What it does |
|---|---|
| Sanctions alert triage | Fine-tunes a small model to decide ESCALATE or CLOSE on screening alerts, gates it on label accuracy and recall, and publishes it. Ships with a synthetic sample of 400 alerts. |
| Chat assistant (SFT → GGUF → Ollama) | Fine-tunes a small chat model on Hub conversations, compares it with the base model, merges it and exports GGUF files for Ollama. |
| Preference tuning (DPO) | Aligns a model on chosen/rejected answer pairs with DPO, then publishes the adapter. |
| Document and image understanding (vision) | Fine-tunes a vision-language model on image, question and answer rows. |
| Reasoning with rewards (GRPO) | Trains a small model on GSM8K math problems with rule-based rewards. |
| Compare base models | Scores two open models on your own test set and on a standard benchmark. |
| CPU smoke test (no GPU needed) | Trains a 135M-parameter model for 20 steps on 60 sample chats, to check your setup in a few minutes. |
| Blank canvas | An empty pipeline. |
24 blocks with typed connections and a settings form for each:
| Group | Blocks |
|---|---|
| Data | Load dataset, Clean and dedupe, Format as chat, Format as preference pairs, Format as vision chat, Train / validation / test split, Merge datasets, Generate synthetic data |
| Model | Base model, Registered model |
| Training | Fine-tune (SFT) with LoRA, QLoRA or full fine-tuning; Preference tuning (DPO / ORPO / SimPO / KTO); Vision fine-tune (SFT); Reinforcement tuning (GRPO) |
| Evaluation | Evaluate on test set (with base-versus-tuned comparison), Benchmarks (lm-eval), Prompt-injection test, Quality gate |
| Export | Merge adapter, Export GGUF, Publish and register |
| Deploy | Deploy endpoint, Ollama package, Endpoint smoke test |
Pipeline editor. Drag-and-drop canvas with type checks on every connection, presets, a YAML view you can edit, validation as you edit, undo and redo, and autosave.
Six places to run. Kaggle (free), Notebook (Colab, Lightning or any Jupyter), My computer, Modal, RunPod and Hugging Face Jobs. Each run has a time limit and, on paid targets, a cost limit.
Hardware estimator. A memory breakdown (weights, adapters, gradients, optimizer, activations) for any model and method, which of 25 GPU types fit, and estimated time and cost per provider, with free options first. Live prices for RunPod, Vast.ai and Hugging Face. The same estimate appears in the Run pipeline dialog.
Model catalog. 237 open models you can fine-tune (Qwen, Llama, Gemma, Mistral, Phi, DeepSeek, Granite, OLMo, SmolLM, InternVL and more) with license, gating and memory needs, plus live Hugging Face Hub search.
Runs. Live status for every step, logs, metric charts, evaluation reports, artifacts, Cancel and Run again, comparison of two to four runs, and caching so unchanged steps reuse earlier results.
Datasets. Upload CSV, JSONL, JSON, Parquet or ZIP files (up to 50 MB), or use a Hugging Face dataset or a URL. Tuneline detects the format and shows a preview, duplicate and personal-data counts, and versions.
Registry. Versioned models with stages (staging, production, archived), model cards and lineage from dataset to run to model to deployment.
Deployments. OpenAI-compatible endpoints on RunPod Serverless, Hugging Face Inference Endpoints or Modal, with
stop, resume and delete, and a ready-to-copy curl example.
Playground. Chat with one or two deployments or Hub models side by side, with latency per answer.
Golden sets. Save good answers from the Playground with Save to golden set, then evaluate models against them with the Load dataset block's Golden set source.
Updates, 28 September 2026¶
- Redesigned interface: an indigo and violet theme, a deep indigo sidebar, a new sign-in page and clearer empty states.
- Sign-in: the Continue with GitHub button now appears only when GitHub sign-in is turned on for the site. Before, it was always shown and failed when the provider was off.
- My computer, one-time setup: Settings → Your computers → Add a computer now offers Download setup file for Windows. Double-click it to install the runner and start a background worker that starts again whenever you sign in to Windows. You no longer need to keep a window open. See Install.
- Live run status on the pipeline canvas: while a run is going, a banner above the canvas shows where it runs, the current step with its message and progress, and how many steps are done. Each step on the canvas is colored by its status. Use Hide status on canvas / Show status on canvas to toggle the colors, and Open run to jump to the run page.
- Clearer messages when your computer is offline: the Run dialog, the run page and the banner say whether your computer is connected, and point you to Settings → Your computers to finish its setup.
- Runner messages now refer to "Settings > Your computers", matching the app.
Updates, 29 September 2026¶
- No more blank pages after an update: a tab opened before a new version of Tuneline Studio was deployed could go blank when you opened another page, until you refreshed. The app now reloads itself once to the new version, landing on the page you opened; if that doesn't help, the page says what happened and offers Reload page while the sidebar keeps working.
- Playground, ready to chat: Model now lists every chat model Hugging Face Inference Providers serve right now, popular ones first, with the number of providers, the context length and the price. A model is picked for you, so you can type and send straight away. The page warns before you type when your Hugging Face token is missing, and offers Deploy for your fine-tuned models (Inference Providers can't run private models). Errors say what to do: a token without the Inference Providers permission, a gated model, or a model no provider serves.
- Hardware page, now actionable: pick one of your pipelines in Estimate for to fill the job from it; pick a GPU under GPUs that fit to see only its offers; Run here next to an option opens the pipeline's Run dialog with that provider and GPU already picked. Your computers shows whether the job fits on each of your own GPUs. The table shows the best 8 options first (Show all for the rest), and only providers Tuneline can start unless you untick Only where Tuneline can run it.
Updates, 30 September 2026¶
- Sign in with GitHub or Google: new accounts are created with Continue with GitHub or Continue with Google, with no email to confirm. Each button appears once the site has that sign-in turned on. Accounts that already have a password still sign in with their email and password, and signing in through GitHub or Google with the same email address opens the same account. Create account and Email me a link are gone: both depended on sending email.
- Sign-in errors explained: when a GitHub or Google sign-in fails, the sign-in page says why, for example an email address that can't create accounts or a sign-in cancelled on GitHub.
- Privacy policy and terms of service: see Privacy policy and Terms of service, also linked from the sign-in page.
Known limitations¶
- Not yet verified against every provider. Launching runs on Modal, RunPod, Hugging Face Jobs and Kaggle, and creating endpoints, are built against the providers' documented APIs and covered by tests with mocked services, but haven't all been checked with real accounts yet. Please report what you find (see Ask a question).
- Kaggle runs can't be cancelled right away. Kaggle's API has no usable cancel, so a cancelled Kaggle run stops at the runner's next check-in, within about 15 seconds while it is running.
- QLoRA on native Windows depends on the
bitsandbytesbuild. LoRA and full fine-tuning work natively; for QLoRA (and Unsloth), WSL2 with Ubuntu is the most reliable option. QLoRA isn't available on macOS. - Unsloth falls back to TRL. Current Unsloth releases require an older TRL and Transformers than Tuneline uses, so the training steps' Use Unsloth setting has no effect for now.
- vLLM evaluation falls back to Transformers in cloud jobs, because the job images use a PyTorch version older than current vLLM releases need. Endpoints use their own vLLM images and aren't affected.
- No multi-GPU data parallel training. With several GPUs, a run spreads the model's layers over them (more memory, no speed-up). The estimator assumes no speed-up from extra GPUs.
- Estimates are approximate. Time estimates assume a fixed efficiency and a fixed start-up time, and will be tuned against real runs.
- No automatic clean-up on workers. The work folder on your computer (
~/.cache/tunelineby default) keeps every step's results and grows with each run. Delete it when you need the space. - First Modal run is slow to start. The first run in a Modal workspace builds the job image, which can take several minutes.