SFT

Supervised fine-tuning trains the model to reproduce known answers. Use it when your dataset already contains the responses you want. An SFT adapter is also the usual starting point for preference optimization.

Configuration

[run]
algorithm = "sft"

[model]
path = "base.gguf"

[output]
path = "sft-adapter.gguf"

[lora]
rank = 8
alpha = 16.0

[training]
ctx = 512
epochs = 3
lr = 0.0001
lr_scheduler = "cosine"
warmup_steps = 10

[sft]
data = "train.jsonl"

retrograd train sft.toml

With chat JSONL, only assistant turns are trained. With plain text, every token is. Rows are shuffled at each epoch (sft.shuffle = false keeps file order). See Datasets.

Tool conversations

A tool record trains each assistant turn - its prose, its call markup in the model's own format, and the end-of-turn token the model stops on - and nothing else: the catalog, the observations and the template's framing are context. This is the stream an agentic rollout samples, so SFT on successful traces is the usual warm-start before agentic GRPO: a policy that cannot yet call a tool validly earns the same zero reward on every member of a group, and GRPO learns nothing from it.

Template variables

[sft]
data = "traces.jsonl"
template_variables = { enable_thinking = false }

Handed to the chat template while the dataset is rendered, like [agent].template_variables during a rollout. A warm-start before an agentic run must use the run's own: under other variables the template writes another format, and the model learns a turn it will not be asked to sample.

Batch size and memory

Key	Default	Meaning
training.ctx	128	Longest sequence, in tokens.
training.micro_batch	32	Tokens per forward/backward pass. The main memory knob.
training.gradient_accumulation	1	Passes accumulated per optimizer step.

micro_batch × gradient_accumulation must divide ctx. If the run runs out of memory, lower micro_batch and raise gradient_accumulation to keep the same step size. More options are in Performance and memory.

Evaluation

[evaluation]
data = "eval.jsonl"
patience = 3

Evaluation runs after every epoch and reports held-out loss and perplexity. patience stops the run after that many evaluations without improvement. After training, compare the base model and the adapter on the same data:

retrograd bench sft.toml --data eval.jsonl --adapter sft-adapter.gguf

Training base weights

By default only a LoRA adapter is trained. training.trainable trains the model's own weights instead, and works the same way for every algorithm:

trainable	Trains	Needs
lora (default)	A LoRA adapter	[lora]
full	Every trainable base tensor	nothing else
partial	The base tensors selected in [trainable]	[trainable]
hybrid	An adapter plus norms and biases	[lora] and [trainable]

[training]
trainable = "partial"

[trainable]
layers = "last:4"
modules = ["attn", "ffn"]
norms = true

[output]
path = "tuned.gguf"
kind = "model"   # a standalone GGUF, usable without Retrograd

Quantized tensors are never trained, so the selected tensors must be stored in F32, F16 or BF16; which of these each backend accepts is in the support matrix. See [trainable] and [output] for every option.