Fine-Tuning

OK

Identifier of the base model to fine-tune. If omitted, the default chat model is used.

Training stage: 'sft' (default) fine-tunes on chat conversations with the chat template applied; 'pretrain' continues pre-training on raw text (assistant-only masking does not apply).

Raw training text for the 'pretrain' stage. Combined with any uploaded dataset file.

Inline training samples as conversations. Each item is a list of role/content messages; assistant turns are supervised. Provide this or dataset_file_id.

File id (from /lmkit/v1/files/upload) of a dataset to train on: JSONL chat, ShareGPT, Alpaca, plain text, or a ZIP archive of these. Format is auto-detected. Provide this or dataset.

LoRA rank (inner dimension of the adapter). Higher adds capacity and size. Default 8.

LoRA alpha scaling factor. Effective scale is alpha/rank. Default 16.

Which modules receive adapters: 'attention' (default), 'attention_and_feedforward', or 'all'.

Number of passes over the training set. Default 3.

Initial AdamW learning rate. Default 1e-4.

Samples accumulated per optimizer step: gradients combine across this many samples before the weights update, multiplying the effective batch size at flat memory. 1 steps on every sample.

Rank-stabilized LoRA: scales the adapter by alpha/sqrt(rank) instead of alpha/rank, keeping higher ranks trainable. The scaling folds into the saved artifact.

Learning-rate schedule: 'cosine' (default), 'constant', 'linear', 'cosine_with_restarts', or 'polynomial'.

AdamW weight decay.

Gradient clipping by global norm before each optimizer step. Default 1.0; 0 disables clipping.

Floor the decaying schedules approach. Ignored by the constant schedule.

Fraction of steps spent warming the learning rate up from zero. Default 0.

Fraction of samples held out for per-epoch validation. Default 0.05.

Compute loss only on assistant tokens (standard for instruction tuning). Default true.

Pack consecutive short samples into shared training windows. Loss never crosses a sample boundary; image samples are never packed. Default false.

Seed for reproducible adapter initialization. 0 means non-deterministic.

Maximum tokens per training sequence (the training window). Samples longer than this are skipped, and every sample is padded to it during packing. 0 (the default) sizes the window to the longest sample.

Save a training checkpoint (adapter weights + optimizer state) every N optimizer steps, enabling resume. 0 disables checkpointing.

Stop the run after this many validation passes without improvement, keeping everything trained so far. Requires a validation split. 0 disables early stopping.

Produce the artifact from the weights at the BEST validation loss instead of the last step: the run snapshots the adapter whenever validation improves. Requires a validation split.

Tokens evaluated per training micro-batch. Smaller values cut activation memory linearly; larger values run faster when memory allows. 0 (the default) picks 512 or 256 to divide the window.

LoRA+ learning-rate ratio: the zero-initialized B matrices train at this multiple of the base learning rate, which speeds convergence at unchanged memory. 16 is the common value; 0 (the default) trains both sides at the base rate.

NEFTune noise alpha: training adds uniform noise scaled by alpha over sqrt(tokens x embedding width) onto the input embeddings, a regularizer that counters overfitting on small datasets. Validation always runs without noise. 5 is the paper's default; 0 (the default) disables it.

First transformer block that receives adapters, inclusive. 0 (the default) starts at the first block. Restricting the range cuts adapter memory and backward compute proportionally.

Last transformer block that receives adapters, inclusive. 0 (the default) extends to the last block.

Train from the model's full-precision (F16/BF16) variant when its repository publishes one, downloading it on first use. Ignored for custom paths and models with no published variant.

For merged-model output: quantize the merged GGUF to this precision (q4_k_m, q5_k_m, or q8_0). Empty keeps the merge at the base's precision. Ignored for adapter output.

Artifact to produce: 'adapter' (default, a small LoRA GGUF) or 'model' (base merged with the adapter).

Accepted

Bad Request

The training job identifier.

OK

Not Found

The training job identifier.

OK

Not Found

The training job identifier.

OK

Not Found