Accelerate Fine-Tuning Without Sacrificing Control
Ertas Studio gives ML engineers a visual fine-tuning workspace on managed cloud GPUs, with parallel runs, a guided dataset builder and export to GGUF or safetensors, so you spend less time on boilerplate and more on model quality.
The Challenges You Face
Boilerplate Slows Down Experimentation
Setting up training loops, data loaders, adapter configs and quantization scripts for every experiment is repetitive work that does not improve model quality. Yet skipping any step risks silent regressions or wasted GPU hours.
Training Data Lives in Scattered Notebooks
Datasets get cleaned and edited in ad-hoc notebooks and spreadsheets with no versioning. When a model regresses, tracing the cause back to a specific change in the data is slow.
GPU Cost Management Is a Full-Time Job
Provisioning the right instance type, managing spot interruptions and tuning batch sizes to keep GPUs busy is infrastructure work that competes with actual research time.
Handoffs Between Domain Experts and Training Are Fragile
Subject-matter experts review examples in one place, ML engineers train in another, and handoffs happen through shared drives with naming conventions that break. There is rarely a single place linking a training run to the dataset version that produced it.
How Ertas Solves This
Ertas Studio gives you a visual layer over the fine-tuning workflow you already understand. You still choose the base model and hyperparameters (learning rate, steps or epochs, batch size, warmup, optimizer, GPU), but instead of writing training scripts you configure them on a canvas that launches LoRA fine-tuning on managed cloud GPUs. Run several configurations in parallel and compare them side by side.
Data Craft, Studio's guided dataset builder, brings structure upstream. Upload JSONL or import a dataset from Hugging Face, then edit and version examples as USER and ASSISTANT turns and track which ones are approved. Domain experts can work on the same dataset without touching code.
Every trained model lands in Hub. Export it as a Q4_K_M GGUF for llama.cpp, Ollama, LM Studio or on-device use, or as full safetensors for vLLM and other server runtimes. The model is yours to run anywhere, with no Ertas dependency at runtime.
Key Features for ML Engineers
Hyperparameter Workspace
Configure learning rate, max steps or epochs, batch size, warmup, gradient accumulation, optimizer and GPU tier through a structured interface. Duplicate a run, change one setting, and compare the two.
Versioned Datasets in Data Craft
Edit examples as USER and ASSISTANT turns, version them and track approval status. Datasets are stored so any run can be repeated, and you can delete them when you are done.
Model Hub and Export
Every trained model is listed in Hub. Download it as GGUF for llama.cpp, Ollama and LM Studio, or as full safetensors for vLLM, and evaluate it on your own prompts with your own tooling.
Managed Cloud Training
Training jobs run on managed T4 and A10G GPUs, several at a time, without provisioning instances. Studio handles drivers, checkpoints and GGUF conversion so you can focus on the experiment.
Why It Works
- On the QMSum meeting-summarization benchmark, a 1.2B model trained on Ertas scored 35.41 ROUGE-1 against 32.43 for the best of five hosted frontier models: a small model trained for one task can beat far larger general models on that task.
- GGUF export produces a Q4_K_M quantized model by default, and full safetensors exports serve on vLLM and other server runtimes.
- The model catalogue spans Llama, Qwen, Gemma, Mistral, Phi and LFM families from 230M to 8B parameters, and compatible Hugging Face models can be imported by URL.
Example Workflow
Your team has a batch of domain documents that need to become training data for a specialized extraction model. A data engineer converts them into JSONL, with the source text as the USER turn and the target fields as the ASSISTANT turn, and uploads the file to Data Craft. A domain expert reviews the examples there, fixes the wrong extractions and approves the set.
The ML engineer attaches the dataset to three runs on the Studio canvas: Qwen 2.5 7B and Llama 3.1 8B at the default learning rate, plus a Qwen run at a lower learning rate. All three train in parallel on A10G GPUs. The engineer downloads the GGUF files, scores each against a held-out prompt set, and promotes the winner, which the platform team serves with vLLM from the safetensors export.
Related Resources
Ship AI that runs on your users' devices.
Free plan with 30 credits/mo, no card required. Paid plans from $10/mo USD.