Validate fit and pin every cost input. No credits move and no GPU starts.
Documentation
From dataset to verified artifact.
A practical guide to the managed Farka workflow: validate the data, build an exact quote, cross the spend boundary deliberately, compare quality, and take the selected artifact.
Start with workspace access
Create a workspace, or accept an invitation to join one that already exists. Every member accepts the current legal documents and works inside an explicitly scoped workspace.
- Upload or import a training dataset and an independent evaluation holdout.
- Select a catalog model and keep the recommended method and automatic accelerator choice for the first estimate.
- Review the exact quote. No compute starts while calculating it.
- Reserve credits and start training.
- Review paired quality evidence before exporting an artifact.
Dataset contract
Structured JSONL, JSON, CSV, GZ, text documents, agent traces, and process specifications are accepted. Paid runs require a separately profiled, non-empty holdout. Invalid or unsafe rows block the quote rather than becoming accelerator input.
{"prompt":"Question or task","completion":"Expected response"}
{"prompt":"Another unseen input","completion":"Another target"}Keep training and evaluation examples disjoint. Farka records row counts, token estimates, bytes, schema, split evidence, and content checksums.
The estimate is a workload contract, not a plan price.
Farka cannot price a run from the model name alone. The estimator combines the profiled training and holdout data, model size, objective, method, epochs, compatible accelerator and current rate evidence, storage, retention, evaluation work, delivery artifact, and software margin.
Buy only the missing balance through Stripe. Credits remain available to the workspace.
Reserve the quoted credits and start training. This is the compute boundary.
- GPU compute
- Estimated aggregate GPU hours multiplied by the pinned compatible accelerator rate.
- Data and retention
- Stored dataset and artifact bytes, plus the retention window selected for this run.
- Evaluation and operations
- Holdout evaluation, retrieval or reward checks, orchestration, and platform operating cost.
- Delivery
- A target-specific fee for direct weights or a runnable deployment bundle. Managed delivery appears only when explicitly enabled.
- Software margin
- Shown in the quote alongside hard costs rather than hidden in an unitemized total.
Your first eligible run is $5. The offer covers supervised LoRA or QLoRA with up to 1,000 rows, 1,024 tokens per row, one epoch, a model up to 2B parameters, one accelerator, adapter-weight delivery, and no retention. The quote also confirms that current serving cost fits the offer. One started trial per workspace; estimates and unlaunched jobs do not consume it. Improving a model later costs less than training a new one, because a continuation pays a reduced delivery fee on a model whose delivery you already bought.
Successful provider usage is reconciled against the reservation. If measured usage exceeds it, available credits cover the difference before artifacts unlock; an unpaid balance keeps delivery locked.
Workspace access is required before an estimate can be calculated.
Run lifecycle
estimate→awaiting credits→queued→running→evaluating→ready
The accepted quote remains attached to the job. Launch reserves credits and provisions compute; failure or cancellation before successful settlement releases the reservation. When training and evaluation finish, Farka emails the verified submitter or workspace owner.
Promotion is stricter than completion
Farka compares the untouched base model and tuned model — or, for a continuation, the predecessor model — on the same unseen holdout. Text promotion requires coherent paired results, valid output, no safety regression, and a task threshold. A regressed run remains visible for review, but delivery stays locked.
Choose delivery before paying
Download open weights or take a runnable Docker bundle. Bundles include the tuned weights, checksum-bound runtime, native /invoke route, OpenAI-compatible API, and TypeScript/Python clients that run in your environment. Weight exports ship the tokenizer and config alongside the safetensors, so they load in Transformers, vLLM, or a GGUF conversion for llama.cpp, LM Studio, and Ollama without our SDK. Farka-managed hosted endpoints are not part of the current launch.
Improve a model you already trained
Open a dataset lineage on the first upload, then add each later version to the same lineage. The lineage freezes the holdout by row content, so new rows never move an existing row across the train/holdout boundary and two iterations are measured on the same terms. Re-uploading content the lineage already holds returns that version instead of creating a new one, so a retry cannot inflate the history.
Train the next iteration from the delivered model rather than the untouched base weights. The run warm-starts from that adapter, and its reported change is measured against the predecessor: it answers whether this round helped. The cumulative change is measured from the first iteration’s baseline, which is the comparison against the base model.
Automate training and delivery
The stable control-plane API exposes dataset intake, quotes, jobs, launch, evidence, usage, and unlocked artifacts. Runnable bundles provide native and OpenAI-compatible inference in infrastructure you operate.