Glossary

Plain-language definitions of the terms used across these guides. If a term is missing, ask — it becomes an FAQ entry.

Term Definition
Adapter Small trained weights (from LoRA/QLoRA) that attach to a base model to change its behavior — no full retrain.
Baseline model DeepSeek v3.2 — the platform's reference model. Other models' credit multipliers are measured against it.
Hosted The default: k3ld runs your training and endpoints on k3ld-managed infrastructure — no AWS account needed.
BYO AWS "Bring your own AWS" — you provide the AWS account and k3ld deploys/operates into it.
Self-hosted on AWS The full platform runs in your own environment, operated by your team.
Canary Route a small % of traffic (e.g. 10%) to a new model variant before full cutover, so regressions surface early.
Catalog The unified view of all your connected data sources (also "metastore").
Checkpoint / checkpoint-resume A saved mid-run state; interrupted runs resume from the last checkpoint instead of restarting.
Context window How many tokens a model can "see" in one call. Locked at base-model choice — you can't extend it later.
Credit multiplier Per-model metering factor = real blended price ÷ baseline price. Cheap models meter below 1.0, pricey ones above.
Credits The platform's usage unit for metered calls (Ask AI, LLM Format, gateway, storage overage). 1 credit ≈ $0.01 of platform cost at the reference rates.
DPO / ORPO / GRPO Post-training methods: DPO/ORPO tune a model to prefer chosen answers over rejected ones; GRPO tunes with a verifiable reward (tests, exact match) — the R1-style reasoning recipe.
Drift When a model's input distribution shifts (the world changed) and its outputs quietly degrade.
Glob pattern A wildcard file pattern (e.g. raw/customers_*.csv) that matches many files — lets one pipeline consume rolling data automatically.
GPU smoke test A short GPU run that validates your data + config before the full run spends — fails in minutes, not hours.
Instance / instance type A specific AWS compute box (e.g. g5.4xlarge = 1× A10G 24GB). Costs are per-hour.
Instance grant Admin-approved set of instance types a tenant/user may deploy or train on.
Keep-for-later Delete an endpoint's billing but keep the model + config for a one-click resume (≈ $0 + small storage).
Lineage The recorded chain of where a table came from and what consumes it.
LoRA / QLoRA Parameter-efficient fine-tuning: adapt ~0.1–1% of the weights instead of retraining the whole model.
Managed S3 The platform's auto-provisioned storage bucket for your uploads and artifacts.
MCP Model Context Protocol — the open standard that lets AI agents call tools/data. The platform exposes your governed data as MCP tools.
Metadata Information about data (schema, counts, lineage) — as opposed to the data rows themselves.
MoE Mixture-of-Experts — an architecture where only a few "expert" parameters activate per token, giving more brain per FLOP.
Multimodal A model that can take non-text input (images, etc.). Locked at base-model choice.
On-demand vs spot On-demand = guaranteed instance at the listed hourly rate. Spot = up to ~60–70% cheaper but can be reclaimed.
Post-training Alignment/hardening after fine-tuning: preference tuning (DPO/ORPO) or verifiable-reward reasoning (GRPO).
Preset A guided fine-tune recipe (chat, instruction, SQL, support) that pre-fills base model, sample data, and hyperparameters.
Rerank Re-order a list of candidate documents by relevance to a query (over embeddings).
Run-confidence cache A record of past run outcomes by config fingerprint — tells you 🟢/🟡/🔴 how likely this exact recipe is to succeed before you spend.
Self-heal Automatically retrying deploy-stage failures (with real error surfacing + admin approve/stop).
Serving artifact / tarball The packaged, verified model the endpoint actually loads — built automatically when you deploy a raw fine-tune.
Tenant / metastore Your org's isolated scope. Records, runs, endpoints, and grants are owned per tenant.
Token A chunk of text a model reads/writes (~¾ of a word in English). Costs and context are measured in tokens.
VRAM Video memory on a GPU. The platform VRAM-checks so you can't create an endpoint a model can't fit.