Skip to content

diffuse-coordinator finetune ​

Fine-tune a model on a file, in one command.

Imports the corpus, chooses hyperparameters from its size and the machine, and starts the run. Does not serve the result: the command for that is printed when it finishes.

Synopsis ​

diffuse-coordinator finetune <MODEL> <FILE> [OPTIONS]

Arguments ​

argumentrequireddescription
MODELyesThe model to adapt, as shown by model list
FILEyesThe corpus: one JSON object per line, each with a messages array

Options ​

flagvaluedefaultdescription
--config<CONFIG>$DIFFUSE_COORDINATOR_CONFIGConfiguration file. Its [admin] section says where to connect
--endpoint<ENDPOINT>$DIFFUSE_COORDINATOR_ENDPOINTCoordinator endpoint, e.g. https://coordinator.internal:7443
--ca-cert<CA_CERT>$DIFFUSE_CA_CERTThe deployment CA certificate (PEM)
--cert<CERT>$DIFFUSE_CERTThis process's certificate chain (PEM)
--key<KEY>$DIFFUSE_KEYThis process's private key (PEM)
--classification<CLASSIFICATION>internalWhat this data is, in your organisation's own words
--as-dataset<DATASET_KEY>-Import the corpus under this name rather than the file's own
--as<ADAPTER_KEY>-What to call the adapter. Derived from the model and the corpus by default
--pool<POOL>-Which pool to place it in. Every healthy node by default
--rank<RANK>-LoRA rank. Higher learns more and costs more
--alpha<ALPHA>-LoRA alpha; the merged delta is scaled by alpha/rank. Twice the rank by default
--targets<TARGETS>-Which projections to adapt, comma separated
--epochs<EPOCHS>-Passes over the corpus. Chosen from its size by default
--learning-rate<LEARNING_RATE>-Step size
--batch<BATCH>-Examples per step
--max-seq-len<MAX_SEQ_LEN>-Tokens per example. Taken from the longest row by default

Notes ​

The one-command path: it imports the corpus, chooses hyperparameters from its size and from the machine, and starts the run. It does not serve the result: model serve <base>+<adapter> does that, once you have looked at the losses.

Examples ​

bash
$ diffuse-coordinator finetune qwen2.5-3b berichte.jsonl
imported 2 412 examples
  rank 16, lr 1e-4, 3 epochs, batch 4, chosen from the corpus and rechner-01
  job 4f2a9c started

Watch it:  diffuse-coordinator job watch 4f2a9c

← All commands

Diffuse Enterprise is commercial software. Diffuse Open is AGPL-3.0.