Skip to main content

Fine-tuning

EcoLink's managed fine-tuning takes you from a dataset to a deployed LoRA adapter without leaving the console. Bring your own training examples or have us synthesize them from seeds + a teacher model; pick a base model and GPU; review the cost; click Approve & Train; deploy the resulting adapter to any compatible model instance.

In other words: distill a digital-employee skill into a small open-source model that runs on EcoLink GPUs, cutting per-call cost vs. commercial LLMs by 10–100× while keeping the workload on one platform end-to-end.

When to use managed fine-tuning vs. alternatives​

You want…Use
Train a LoRA adapter, deploy it as an OpenAI-compatible API, no infra to manageManaged Fine-tuning (this guide)
Full control over the training loop (custom loss, custom dataloader, full-weight fine-tune)Manual fine-tuning on a GPU instance
Use an off-the-shelf platform model (no training)Platform Models
Deploy a model you already trained elsewhereUser Inference

End-to-end flow​

┌──────────────┐    ┌──────────────────────┐    ┌────────────────────┐    ┌──────────────────────┐
│  /datasets   │───▶│  /datasets/{id}      │───▶│  /fine-tune/new    │───▶│  /fine-tuned-models  │
│  upload or   │    │  validate; (seeds)   │    │  pick base, GPUs,  │    │  attach adapter to a │
│  enter seeds │    │  generate synthesis  │    │  hyperparams, cost │    │  LoRA-enabled host   │
└──────────────┘    └──────────────────────┘    └────────────────────┘    └──────────────────────┘

Each box is a console page; arrows are clicks. The whole loop takes ~10 min for a typical 1 000-example BYO upload on a 7B base, longer if you go through teacher synthesis or pick a 30B+ base.

  1. Prepare a dataset (/datasets). Two paths:
    • Upload a JSONL of chat-format training examples you already have. → Datasets.
    • Upload a JSONL of seeds + a skill prompt; we'll fan each seed through a teacher model and produce trainable rows. → Synthesis.
  2. Launch a fine-tuning job (/fine-tune/new). Pick a base model, GPU type/count, LoRA hyperparameters. Review the cost estimate. Click Approve & Train. → Launching a job.
  3. Deploy the resulting adapter (/fine-tuned-models). Attach to an existing LoRA-enabled model instance, or launch a fresh one. The adapter loads at runtime — no pod restart, no extra GPU. → Deploying an adapter.

Supported base models​

Any LLM the platform marks as "fine-tune eligible". The wizard's base-model dropdown only lists bases that are eligible and have an active deployment in the region you're targeting; what you see is what you can train against. The dropdown updates as the platform adds new bases.

Vision, image, audio, and video models are not supported in this release.

What "LoRA only" means​

Managed fine-tuning trains a LoRA adapter — a small set of low-rank matrices that modify the base model's behavior. Adapters are typically 100–500 MB; the base model stays unchanged. At serve time, vLLM loads the adapter alongside the base in GPU memory and applies it for requests that target this adapter's alias.

Why LoRA, not full-weight fine-tune:

  • 50–100× cheaper to store (a 7B full-weight is ~14 GB; a 7B LoRA is ~150 MB)
  • Multiple adapters share one base in memory — N skills, one GPU
  • Fast to attach/detach at runtime (no pod restart)
  • Quality matches full-weight for the kind of skill-tuning most teams need

If you have a use case that needs full fine-tuning (e.g. you're trying to teach the base entirely new factual knowledge), use the manual workflow — you have full control over the training loop on a GPU instance.

What gets billed​

  • Datasets: free to upload; free to validate.
  • Synthesis (teacher-gen path): free on EcoLink — you pay your teacher API provider directly via the API key you supply (BYOK).
  • Training: per-minute GPU meter, no upfront hold. Balance gate at submit and resume time. Auto-pause if balance hits zero mid-train — top up and click Resume to continue from the latest checkpoint. Each per-minute charge lands as a row on /billing.
  • Serving: standard inference-instance billing (per-GPU-hour) for the LoRA-enabled host. Adapter attach/detach is free.

Limits​

  • Datasets up to 50 MB per file.
  • At least 100 examples to start training (1 000+ recommended for usable quality).
  • Up to 16 adapters attached to one model instance simultaneously (configurable per instance).
  • Adapter file lives on EcoLink for the lifetime of the account; no expiry, can be detached/deleted.

Next steps​