Skip to content
CRUCIBLE

Launch

New run

Four fields instead of twelve CLI steps. Crucible validates the dataset before anything is uploaded, funds the fine-tuning sub-account rather than the inference one, and takes over the acknowledgement deadline the moment the task is delivered.

01 · Dataset

chat-messages · instruction · text-completion

Drop a .jsonl file, or

JSONL · UTF-8 · one format throughout · at least 10 examples · up to 8.00 MB

or paste JSONL directly

02 · Target

Network
Base model

Qwen3-32B is mainnet-only.

03 · Training configuration

0G accepts exactly these five parameters and rejects a config with any extra or missing key — after the task is funded. These are the values from 0G’s own working example, not the docs’ template, which differs.

Embedding noise during training. Higher can improve instruction-following on small datasets.

Passes over the dataset. Multiplies the training fee linearly.

1–4. Drop to 1 if the run hits out-of-memory on the H200.

0.00001–0.001. Decimal notation only — 0G rejects 2e-4.

Hard cap on optimiser steps. -1 means "run the full epochs".