Know exactly what enters every run
Annotate images or text, inspect class balance and geometry warnings, then freeze a version. Every training run points back to the data it actually consumed.
Annotations · EDA · immutable versions
Open source · Apache 2.0
Data in. Models out. Nothing leaves your machine.
Orinth is an open-source studio for computer vision, NLP, and LLMs. Label, train, test, and serve — all in one app, on your own hardware.
Pronounced OR-inth /ˈɔːrɪnθ/. Runs on your own hardware — no account, and no data sent anywhere by default.
One install, no sign-upCV · NLP · LLM in one appDocuments → datasets with Data RecipesApache 2.0, foreverCloud is optional, never required
Data Recipes
Your documents are already knowledge. Make them trainable.
Turn PDFs, DOCX, Markdown, CSV, JSONL, and plain text into instruction pairs, question-answer records, or classification data. Review every generated row before it enters Dataset Studio.
Deterministic rules are the local default. LLM-assisted generation is opt-in, and every committed record keeps its source provenance—so the path from model output back to the original document stays visible.
Read the Data Recipes guide01
Add sources
PDF · DOCX · MD · CSV · JSONL
02
Generate
Rules by default · LLM optional
03
Review
Edit, approve, or exclude each row
04
Commit
Versioned dataset with provenance
Live demo
Seven connected surfaces, running in this page.
The actual app UI — same layouts, same controls, same task-aware behavior — with the backend swapped for captured state. Build records from documents, draw an annotation, shape a model, replay training, compare it, run inference, then chat with it. Every interaction is live, but the compute is not.
Datasets
Support knowledge recipe
Instruction recipe · documents become reviewable training records
Source documents
PDF, DOCX, TXT, Markdown, CSV, or JSONL. Extraction stays local.
support-handbook.pdf184k chars · 42 pagesPDFready
release-notes.md31k charsMarkdownready
billing-policies.docx76k chars · 18 pagesDOCXready
Turn PDFs, DOCX, Markdown, CSV, or JSONL into reviewable instruction records. Sources, generation, review, and commit stay traceable — with deterministic local fallback by default.
Built for the work between the buttons
Less pipeline glue. More confidence in every handoff.
Each surface removes a place where context usually gets lost—from the exact dataset a run used to the individual prediction behind a test score.
Design the network before you spend compute
Build a model as a visual graph, inspect each node, catch invalid connections, and emit Keras or PyTorch without translating a whiteboard into boilerplate.
Visual graph · validation · code export
Watch the run, not a loading spinner
Follow epoch progress, live metrics, charts, and logs in one view. Replay a captured run to understand its trajectory before you compare or promote it.
Live metrics · logs · checkpoints
See the records behind the score
Metrics reshape around the task, model comparisons stay side by side, and every aggregate opens into the individual predictions that produced it.
Task-aware metrics · per-item review
Use one review surface for every modality
Run an image or text item, tune only the controls that matter for its task, and keep confidence, overlays, timings, and prediction history together.
Adaptive controls · normalized output
Turn a fine-tuned model into a usable tool
Export to GGUF, serve it on your machine, and chat with sampler controls, visible thinking, and optional web search with inline citations.
Private serving · search · citations
Every task
Seven tasks, three modalities, one workspace.
Not a vision tool that also does text. Every task below shares the same projects, the same Dataset Studio, the same training and testing surfaces — and each is scored on the metrics that actually fit it.
Computer vision
3 tasksClassification
One label per image. Bulk relabeling, class balance in EDA.
Accuracy · Macro F1 · MCC · AUC
Object detection
Boxes around objects. YOLO and COCO import, box editor.
mAP@50 · Precision · Recall
Segmentation
Pixel and polygon masks, drawn in the annotation editor.
Pixel Dice · IoU · mAP@50-95
Natural language
3 tasksText classification
One label per record. Same splits, versions, and EDA as images.
Accuracy · Macro F1 · Weighted F1
Summarization
Source text with a reference summary, compared side by side.
ROUGE-1 · ROUGE-2 · ROUGE-L
Question answering
Context, question, and answer span, with per-row inspection.
Exact match · F1
Language models
1 task, 4 methodsInstruction fine-tuning
LoRA, QLoRA, full fine-tune, or continued pretraining.
Loss · Perplexity · Token accuracy
Export & serve
Adapter or checkpoint to GGUF, re-quantize, serve locally.
llm_adapter · llm_hf · llm_gguf
Chat
Sampler controls, a reasoning view, and web search with citations.
tok/s · time-to-first-token
One project can hold datasets from all three. A run records the dataset version it consumed, whichever modality it came from.
Why Orinth
Every competitor owns one slice. This one covers the whole lifecycle.
Local by default
Your data never leaves your machine
One .dmg, no account, no telemetry by default. Datasets, weights, predictions, and logs live in local storage you can point at, back up, or delete.
One install · no sign-up
Every modality
Vision, text, and language models in one workspace
Segmentation, detection, and classification for images. Classification, summarization, and question answering for text. LoRA and QLoRA fine-tuning for LLMs. One project holds all three.
CV + NLP + LLM
Every step
Raw folder to served model without leaving the app
Dataset Studio, Training, Testing, Inference, and Chat are one continuous surface. An empty dataset points at upload; a finished run points at testing; a scored model points at inference.
Label → train → test → serve
The workflow
Six stages, and each one hands off to the next.
Continuity is the organizing idea. You are never exporting a folder to hand to the next tool, because there is no next tool.
- 01
Organize
Projects
A project scopes its datasets, its runs, and the models promoted out of them, so image work and text work never bleed into each other.
- 02
Prepare
Dataset Studio
Upload, label, annotate, split, preprocess, and version without rewriting the originals. Import from the Hugging Face Hub, or build a dataset from your own documents with Data Recipes.
- 03
Train
Training
Start a task-compatible run from a frozen dataset version. Runs execute as subprocesses and stream epoch progress and logs, so a long fine-tune never blocks the API.
- 04
Test
Testing
Score against a held-out split with task-aware metrics, stack models side by side in one comparison run, then read the per-item rows behind each number.
- 05
Inspect
Inference
Run a single item and read the output with its overlays, class confidence, timings, and full run history. Every prediction is logged with its payload.
- 06
Serve & chat
Chat
Export a fine-tuned model to GGUF, serve it locally, and hold a conversation with sampler controls, an optional system prompt, web search with citations, and a reasoning view.
Where it sits
For people tired of stitching six tools together.
Each of these is good at what it does. None of them covers three modalities across the full lifecycle, on your own machine, with a GUI.
| Tool | Modalities | Lifecycle | Runs where | Interface |
|---|---|---|---|---|
| Roboflow | CV only | Data → train | Cloud | GUI |
| Label Studio | Multi | Label only | Self-host | GUI |
| Weights & Biases | Multi | Tracking only | Cloud | GUI |
| Axolotl / Unsloth | LLM only | Fine-tune only | Local | CLI |
| LM Studio | LLM only | Chat only | Local | GUI |
| Orinth | CV + NLP + LLM | Label → train → test → serve | Local, cloud optional | GUI + CLI |
Comparison reflects each product’s primary, documented scope as of 2026. All are actively developed and may have moved.
The name
Orinth reads as a place. It carries orient — to find your bearings — and the -inth of labyrinth, the path through complexity.
Orinth is where you find your way from a folder of raw data to a model that works. Pronounced OR-inth /ˈɔːrɪnθ/.
What it runs
Four model families, one normalized response.
Inference output is normalized across families, so the review surface keeps its shape when the model underneath changes.
| Family | Tasks | Notes |
|---|---|---|
| Ultralytics YOLO | Detection, instance segmentation | YOLO11 and YOLO26 families. Default confidence 0.65, inference IoU 0.7. |
| Keras Applications | Segmentation, classification | MobileNetV2, EfficientNet B0–B7, ResNet50, U-Net + Inception. Trained and served locally. |
| Hugging Face Transformers | Text classification, summarization, question answering | BERT, BART, and friends, fine-tuned locally. Gated repos need a HUGGINGFACE_HUB_TOKEN. |
| LLMs (Transformers + GGUF) | Instruction fine-tuning, serving, chat | LoRA, QLoRA, full fine-tune, or continued pretraining. Export to GGUF to serve and chat locally. |
Language models
Fine-tune an LLM, then chat with it.
Pick a Hugging Face base and a method — LoRA, QLoRA, full fine-tune, or continued pretraining. Orinth detects your accelerator (CUDA, Apple MPS, or CPU) and picks a matching backend, so the configuration you see is the one that runs.
Export the result to GGUF, serve it locally, and open a chat with sampler controls, a system prompt, web search with inline citations, and a view into the model’s reasoning. Completion tokens, tokens per second, and time-to-first-token are reported under every reply.
Read the LLM guide| Method | What it does |
|---|---|
| LoRA adapter | Trains a small tuned delta over the base. Light, and the recommended default. |
| QLoRA (4-bit) | Lowest VRAM; needs CUDA. On MPS or CPU it trains as plain LoRA. |
| Full fine-tune | Updates every weight. Heavy — prefer LoRA or QLoRA on a Mac or CPU. |
| Continued pretraining | LoRA over full text, for domain adaptation. |
More in the workspace
Everything the six stages lean on.
Hugging Face Hub import
Browse the Hub from the catalog and pull a dataset straight into a project, ready to annotate, split, and train.
Data Recipes
Turn your own documents into an instruction dataset — deterministic by default, LLM-assisted when you supply a key.
Architecture Studio
A visual graph editor for model architectures, with Keras and PyTorch emitters.
Model Library
Trained, reference, and uploaded models in one catalog — checkpoints, GGUF files, and LoRA adapters side by side.
EDA and versioning
Read split counts, class balance, and geometry warnings, then freeze a materialized version for a reproducible run.
Guided tours
A first-run walkthrough plus a per-page tour on every surface, so the workflow explains itself.
Editions
Open source at the core. Cloud only if you want it.
The studio is Apache 2.0 and complete on its own. Cloud and Enterprise add hosted GPUs, teams, and the controls a regulated environment needs — they never gate the local app.
Open source
Orinth
The full studio, Apache 2.0. Every modality, every stage, on your hardware.
View the sourcemacOS .dmg
Orinth Desktop
One universal build for Apple silicon and Intel. Drag to Applications and open.
DownloadHosted
Orinth Cloud
GPU-backed runs and team workspaces for when the local machine runs out.
See pricingSelf-hosted
Orinth Enterprise
SSO/SAML, RBAC, audit logs, and air-gapped deployment for regulated environments.
Talk to usGet started
Two commands to a running stack.
A FastAPI and SQLite backend, a Next.js frontend, and make targets that start both together with interleaved logs. Or skip all of it and open the macOS app.
git clone https://github.com/L007/orinth.git
cd orinth
# verify uv, pnpm, Python 3.11, Node
make doctor
# backend :8000 + frontend :3000
make devStart with a project.
Point it at a folder of images or text, and the workspace carries you through to a scored model you can serve.
Orinth is a research and engineering instrument for measuring model behavior. It produces model outputs for human review — it does not provide clinical diagnosis or autonomous decision-making.