TuningSpork

About this app

Windows 0.1.43 makes setup and in-app updates reliable when unrelated programs such as terminal windows are open, and keeps updates working when the Marketplace revises its review policy, with the same security checks. It keeps everything in 0.1.42: longer chats that summarize older turns while preserving the full transcript, per-chat character and custom instructions, and automatic answer length. Studio and Sidekick share a searchable model library with device, source, model type, task and runtime-state filters, result counts, reset and sorting. Downloaded models are distinguished from the exact currently connected chat model; library inventory checks remain quiet and do not imply GPU readiness. The shared model surface offers Chat models and Speech models, with clear transcription and speaking roles and readable names for trained voices. Sidekick keeps dictation separate from an explicitly started voice conversation. Choose the local microphone and speaking voice; built-in Whisper handles transcription, with overrides in advanced settings, review the recognized text, and use spoken replies without a paid inference API. Voice sessions show recording, mute and end controls, stop after configured silence, and rearm only within the enabled session. These controls do not establish voice likeness or model quality. Sidekick's chat screen puts the conversation first: a compact sidebar for chats, models and files, a quiet conversation header, a centered reading column and one composer with model, microphone, voice and permission controls. On that screen, secondary tools stay in a Tools menu, and a small download indicator opens retained transfer activity. Assistant answers retain formatted Markdown and inline source previews. The chat composer keeps a stable position while shared voice models are checked at startup. Overview cards share consistent heading, detail and action rows at the supported window widths. Studio and the other pages retain their existing appearance beyond these requested alignment fixes. Training activity keeps elapsed time and phase visible while the worker is alive; a percentage appears only when work has a known denominator. Voice controls and card layouts stay within their available space, and floating menus remain above page content. Sidekick offers an optional local agent mode for typed or locally transcribed requests. Access starts off; users explicitly select folders and optionally one Windows app, review the permission scope and enable the session. The selected local model can search permitted folders, read bounded text and propose new Markdown, text or Word (.docx) documents. Every document save shows its destination and authored text for confirmation and never silently overwrites an existing file. Word output supports headings and plain paragraphs. Saved documents retain an inline authored-source preview, including newly created Word documents; requested text/code previews and verified local PNG previews also appear in chat. Arbitrary existing DOCX and PDF binary contents cannot be read or rendered by the agent tools. Windows app tools inspect accessible controls and recheck the selected process and observation before clicking or replacing editable text. App actions require review by default; users can explicitly permit supported actions for the selected app during this session. Stop cancels the bounded loop and Revoke disables access. Unsupported controls fail without an arbitrary command, coordinate or clipboard fallback. Model planning ability and context limits affect task completion. No paid inference service is required. Studio and Sidekick share local speech models. Choose a microphone, record explicitly, stop or cancel, transcribe locally into editable text and play generated speech. Sidekick can read replies using the selected voice. Studio Voice lab prepares paired recordings and exact transcripts for genuine speech training, then supports listening comparisons on new phrases. Speech training changes voice and delivery; a conversation model chooses the words. Trained voices remain experimental until the user judges their quality. Speech uses an optional isolated runtime and downloaded models on the selected data volume, and runs on this computer. Model downloads have a dedicated activity view that remains available as you move between pages. It identifies the model and destination, shows transfer bytes and percentage when a total is known, and keeps completion or failure visible. Search and selection controls have larger type, padding and clearer labels. Busy searches, checks and model operations show visible activity indicators. Find a starting model in Studio or Sidekick, check its fit on this computer or a connected Spark, and choose what to do next. Search supported Hugging Face models and original publisher weights, with recommendations, device fit estimates and discovery labels for community models. Fine-tune supported safetensors models in Studio through Choose model, Add examples and Train & compare. The shared collection exposes Chat, Try decisions, Match data fields and Score a record. Downloads and checked Studio exports are visible in both tools; completed experimental Laya checkpoints can be inspected without pretending they passed quality approval. Local Ollama models are available for chat and are labeled chat-only; their quantized files are not fine-tuning inputs. Saved Spark connections are shared between Studio and Sidekick. Password sign-in provisions an app-owned SSH key for later passwordless reconnection; existing OpenSSH aliases and keys remain supported. Saved sign-ins show their key availability and can be refreshed, reconnected, edited or disconnected. Sidebar chat deletion uses a compact confirmation, removes the chat immediately, and retains Undo. Bring labeled data, adapt a supported model in Studio, compare it with its starting model, and use eligible models in Sidekick. Studio provides guided Qwen LoRA experiments, Laya metadata decisions and numeric XGBoost classification. Separate development, target and retention checks expose regressions; completing training does not approve a model. Sharing unlocks only when the supplied checks pass, and those checks do not establish accuracy on other data. Connect Spark guides users through finding or entering their machine, reviewing its SSH host key, and authenticating. Existing OpenSSH aliases and configured keys remain supported. The app finds the remote account's TuningSpork data location and existing runtime automatically; a network drive or manually selected remote mount folder is not required. Saved connections expose Edit and Disconnect in Studio and Sidekick. The Select model picker lists eligible local models and models on the connected Spark, with the device named on each card. Mount & chat loads the selected model on its device; it does not mount a filesystem or transfer model weights. Search by model name, base model or location, filter by device and model type, and sort names. Prompts, included history and retrieved document passages travel to the selected Spark for inference. Compatible metadata and numeric classifiers use their own workflows. Ordinary Chat does not give models shell commands or account access; separately enabled local agent tools use explicit permission scopes. Delete chat requires confirmation; Undo restores its retained local recovery copy. Windows x64 and DGX Spark Linux ARM64 have separate platform downloads under this single TuningSpork listing. Each download bundles Launcher, Studio and Sidekick. Windows setup offers component selection. The Spark AppImage opens Launcher and prepares the bundled per-user suite. Add a missing tool later from Launcher using the verified offline cache. Models, chats, datasets and settings live outside application release folders. App updates checks the DeloresAI accepted catalog. Install update downloads and verifies the approved suite artifact directly, closes the Windows suite automatically and preserves component selection and user data. Current platform availability and acceptance are established by each release receipt, not by this preparation document. This is an experimental preview. Earlier Qwen checks retained a critical error, and independently written Laya cases showed poor transfer; those candidates remained rejected. No model checkpoint is bundled and no general model-quality improvement is promised. Download or import an eligible model and supply suitable data. Models have separate licenses. NVIDIA hardware, a working driver and 64-bit CPython 3.12 are required for model computation on the selected device. Prepare runtime installs exact hash-locked dependencies into a fresh isolated environment only when requested. Initial runtime and optional model downloads need internet; prepared local workflows need no cloud AI service. Both desktop builds can run Studio jobs on a configured Spark over SSH. Remote jobs and chat require a compatible prepared runtime on that host. The native Spark target is ARM64 on Ubuntu 24.04-based DGX OS with NVIDIA GB10 and driver 580 or newer. Spark runtime preparation retains the known NVIDIA cuSPARSELt packaging diagnostic, verifies the original vendor bytes and independently checks CUDA calculations; other dependency errors stop setup. The platform minimum disk budget covers the application install and update, not optional ML runtimes or model checkpoints. Runtime preparation checks its destination, cache and temporary storage separately; models need additional capacity at their selected location. Starting from a downloaded language model selects the model fine-tuning workflow with a valid checkpoint metric; unsupported retained metrics are shown explicitly. The shared model library keeps a permanent 32px activity slot beside its filters, with the refresh or verification phase available on hover. Library and chat model chooser controls/cards keep their position through recurring inventory checks. The chat model chooser stacks its labeled selectors and bounds them to the drawer width so they remain readable without extending beneath the conversation. Sidekick retains draft and model-search focus while background status panels change. Studio Voice lab accepts several CSV or JSONL teaching files together with saved recordings; each file's relative audio paths resolve against that file. Explicit splits remain intact and cross-file duplicates or group leakage are rejected. Qwen language training and local replies use a consistent final-answer template; returned completions are graded unchanged. Speech training aligns both primary and residual codec targets without an extra causal shift. These corrections do not certify voice likeness, personality quality or checkpoint approval.

External services & costs

NVIDIA CUDA through isolated Python workers: No application subscription or developer credits. Hardware, storage and electricity are supplied by the user. 64-bit CPython 3.12 and the driver are prerequisites. PyPI and PyTorch package repositories: The fixed Windows wheel download set is 2,974,104,140 bytes; extraction and the installed runtime need additional disk space. No package-index dependency fallback or source builds are permitted. Hugging Face model downloads: Model-specific licenses, access conditions, storage and bandwidth apply. No checkpoint or paid model entitlement is included. User-controlled SSH compute: The user supplies the remote machine, compatible runtime and network access. TuningSpork does not provide a hosted training service. DeloresAI coordinated suite updates: A Marketplace account and internet are required. The suite remembers its protected desktop sign-in; no copied approval code is required for each update. No developer credits or application subscription. Ollama: Optional local service and model downloads; separate model licenses apply. Windows UI Automation: Uses the user's Windows apps and local model. No paid API or remote desktop service.

NVIDIA CUDA through isolated Python workers — Required
Model computation requires NVIDIA hardware, drivers and a compatible CPython 3.12/CUDA worker environment. Prepare runtime installs dependencies only when requested.
No application subscription or developer credits. Hardware, storage and electricity are supplied by the user. 64-bit CPython 3.12 and the driver are prerequisites.

PyPI and PyTorch package repositories — Optional
User-initiated runtime preparation installs 42 reviewed Windows CPython 3.12 wheels from exact official HTTPS URLs with required SHA-256 verification. An existing compatible runtime may be selected instead.
The fixed Windows wheel download set is 2,974,104,140 bytes; extraction and the installed runtime need additional disk space. No package-index dependency fallback or source builds are permitted.

Hugging Face model downloads — Optional
Download an explicitly selected supported base model; local compatible model files can be used instead.
Model-specific licenses, access conditions, storage and bandwidth apply. No checkpoint or paid model entitlement is included.

User-controlled SSH compute — Optional
Find or enter a Spark, review the SSH host key and authenticate, discover its existing TuningSpork environment, run selected jobs and use an explicitly mounted remote model. Existing OpenSSH aliases/key configuration remain supported.
The user supplies the remote machine, compatible runtime and network access. TuningSpork does not provide a hosted training service.

DeloresAI coordinated suite updates — Optional
Check the accepted platform catalog, directly download and verify the approved artifact, then install the suite after the user chooses Install update.
A Marketplace account and internet are required. The suite remembers its protected desktop sign-in; no copied approval code is required for each update. No developer credits or application subscription.

Ollama — Optional
Chat with user-selected local models through the running localhost Ollama service; quantized Ollama files are not fine-tuning inputs.
Optional local service and model downloads; separate model licenses apply.

Windows UI Automation — Optional
Opt-in selected-app control via native accessible controls, bound to the observed window and process lifetime; inaccessible canvas controls remain unsupported.
Uses the user's Windows apps and local model. No paid API or remote desktop service.

Privacy & permissions

Local training, local chat and local classifier inference do not automatically upload datasets, models or conversations. Selecting Studio SSH compute copies selected job data and workers to the configured host and retrieves results. Selecting a Spark chat model sends prompts, included conversation history and retrieved document passages to that host for inference; model files remain on their selected device. Model/runtime downloads contact their repositories. Update checks contact deloresai.com with normal request metadata and platform/app information. Updates download the exact approved artifact directly from the Marketplace using the saved Marketplace desktop session. Explicit sign-in sends credentials only to the official Marketplace authentication service; passwords are not saved. Sharing a .tspmodel file is a separate user action; reports may contain evaluation examples. Microphone recordings, local transcription and local speech generation stay on this computer; speech operations do not use the selected Spark target. Opt-in local agent file snippets and selected-app observations are supplied to the selected local model on this computer. App-control actions can change the chosen application; Sidekick does not claim that permission prevents model mistakes.

Filesystem: Read user-selected training and evaluation datasets, compatible model folders, imported packages and explicitly attached text documents, Read and write TuningSpork jobs, models, chats and retained deleted-chat recovery files, reports, settings and selected document indexes, Read and write the local Hugging Face model cache and an explicitly prepared isolated Python runtime, Read packaged workers; stage verified update artifacts and install per-user suite components and Launcher cache while preserving separate user data, Copy selected job/worker files to and from a configured SSH host; discover and verify remote model files and start app-owned inference workers there, On explicit Spark setup, create owner-restricted SSH identity/config files and append only the public key to remote authorized_keys; preserve unrelated SSH files and keys., Use the selected microphone only after Record; write local WAV clips/transcripts, retained teaching audio and speech checkpoints, and play selected local audio., With explicit consent: search/read selected folders; preview text/code/PNG; review/save new Markdown, text or Word files; control a selected Windows app. Existing DOCX/PDF contents are unsupported.. Network domains: pypi.org, files.pythonhosted.org, download.pytorch.org, download-r2.pytorch.org, huggingface.co, us.aws.cdn.hf.co, deloresai.com. No OS elevation declared.