config-reference.md

Configuration Reference

中文 | Documentation index

Final configuration is defaults merged with YAML and --set, then validated before data access. Use show-config to inspect effective values.

data

Field	Default	Constraint and meaning
name	ag_news	ag_news or headered generic_csv
data_dir	data/raw	AG News root or directory containing generic train/test CSV
manifest_dir	data/manifests	Prepared CSV, dataset.json, and audit directory
tokenizer	simple_word	Built-in word tokenizer only
text_column	text	Generic CSV text column
label_column	label	Generic CSV label column; must differ from text
vocab_size	30000	Includes four special tokens; at least 4
min_frequency	2	Positive training-vocabulary frequency
max_length	128	Includes BOS/EOS; at least 2
valid_ratio	0.1	Strictly between zero and one, per class
num_workers	0	Non-negative DataLoader workers
max_*_samples	null	Per-split debug limit; null or positive integer

model

Field	Default	Constraint and meaning
name	embedding_bag	embedding_bag, text_cnn, or bilstm
embedding_dim	128	Positive token-vector width
hidden_dim	128	CNN channels or LSTM units per direction
dropout	0.2	[0,1)
kernel_sizes	[3,4,5]	Positive TextCNN list, no larger than max length
num_layers	2	Positive BiLSTM layer count
bidirectional	true	Whether BiLSTM is bidirectional

train and run

Field	Default	Constraint and meaning
epochs	2	Total epochs; target total when resuming
batch_size	32	Positive integer
lr	0.001	Positive number
weight_decay	0.0001	Non-negative number
optimizer	adamw	adamw, adam, or sgd
momentum	0.9	SGD momentum; ignored by other optimizers
seed	42	Non-negative integer
amp	false	Automatic mixed precision on CUDA only
deterministic	false	Request deterministic algorithms
grad_clip	0.0	Clip gradient norm when greater than zero
best_metric	macro_f1	Accuracy or one of four macro metrics
device	auto	auto, cpu, mps, cuda, or cuda:N
output_dir	artifacts	Run root
run_name	null	Single path component; null creates timestamp name

uv run text-classify show-config --config configs/reference_textcnn.yaml \
  --set train.epochs=12 --set train.optimizer=adamw