Fine-Tuning

Stage: Build → extension Fine-tune · What it does: adapt an open base model to your domain, brand, and task with guided presets — no ML specialist required. Uses parameter-efficient methods (LoRA/QLoRA) so you tune ~0.1–1% of the weights: fast, cheap, and portable.

Why it matters

• Own the behavior. The patterns your data teaches the model are yours — competitors calling the same public endpoint can't replicate them.
• Predictable cost. Fixed infrastructure cost, with an estimate and an optional GPU smoke test before every run.
• No data migration. You connect your training source — no export or upload step. With BYO / self-hosted the run stays in your environment; hosted runs on k3ld-managed compute.

Data formats

Bundled example datasets ship with each preset so you can validate end-to-end before bringing your own data.

How to use

1. Build → extension Fine-tune.
2. Pick a use-case preset — it auto-fills base model, a matching sample dataset, and sensible hyperparameters.
3. Swap in your data (JSONL) if you have it.
4. Review the cost estimate and the base-model contract (loader, VRAM, recommended instance, sequence length).
5. Optional: tick "Run GPU smoke before launch" — a short GPU run validates your data + config before the full job spends.
6. Launch. Watch progress; completed runs show a confidence indicator from similar past runs.

Terminal users: the same flow is available via CLI.

Real-world examples

🏦 Financial services — a compliance-aware chat assistant

A bank fine-tunes a chat assistant on approved call-center dialogues (scrubbed of PII in-place). Using the chat (messages) preset with their conversation logs, they get a support model that answers in the bank's tone and policy language — running hosted, BYO AWS, or self-hosted on AWS, so regulators see no data leaving the tenant. Each run gets a fixed cost line, not a surprise token bill.

🏥 Healthcare — clinical summarization, PHI in place

A health system fine-tunes a summarization model on clinical handoff notes stored in their own bucket. LoRA/QLoRA keeps the run on a single GPU instance; the model learns their note structure and abbreviations. In BYO / self-hosted mode the data never leaves the tenant, so the compliance review is short.

🛒 Retail — product descriptions in brand voice

A retailer fine-tunes on thousands of past product listings (instruction: attributes → output: description). The model then generates draft descriptions for new SKUs in the same brand voice and SEO structure — deployed as their own endpoint, so every generated description is their asset.

🧑‍💼 SaaS — tiered support deflection

A SaaS company fine-tunes on historical tickets + resolutions (support assistant preset). The result deflects ~40% of L1 tickets with answers in the product's voice, served from their own endpoint — no per-ticket API fees at volume.

⚖️ Legal — clause extraction

A law firm fine-tunes on labeled contract clauses (instruction/output). Extraction quality on their contract style jumps vs. a general model, and the tuned weights are an in-house asset they can export and keep.

Gotchas

• Chat data must be {"messages":[...]} for the chat format — the platform validates this and tells you exactly what's missing before you spend.
• Base-model choice locks context length and multimodal ability; style is fixable later, context is not. Pick the base by the axis you can't change.
• Small models are the cheapest way to validate your dataset first — fine-tune a tiny model on a sample, then scale up.
• Completed runs can be deployed directly — the platform builds the serving export automatically behind the scenes (no manual step).

Related

• Post-training
• Ship & deploy
• Model catalog
• Cost control