mirror of https://github.com/razor-ai/soup.git
45 lines
1.1 KiB
YAML
45 lines
1.1 KiB
YAML
# SFT Basic Example
|
|
# Fine-tune TinyLlama-1.1B on instruction-following data.
|
|
#
|
|
# The bundled dataset is a 10-row FORMAT sample, not a training set. This
|
|
# config runs end-to-end so you can verify your setup, but it will not
|
|
# produce a usefully fine-tuned model — point `data.train` at your own data
|
|
# for that. See examples/data/README.md.
|
|
#
|
|
# Usage:
|
|
# soup train --config examples/configs/sft_basic.yaml
|
|
|
|
base: TinyLlama/TinyLlama-1.1B-Chat-v1.0
|
|
task: sft
|
|
backend: transformers
|
|
|
|
data:
|
|
train: examples/data/alpaca_tiny.jsonl
|
|
format: alpaca
|
|
# 512: the longest row in any bundled fixture is ~233 tokens, so nothing is
|
|
# truncated and there is 2x headroom. A larger value only inflates the
|
|
# activation memory the VRAM pre-flight budgets for. Raise it when you point
|
|
# this config at longer data of your own.
|
|
max_length: 512
|
|
|
|
training:
|
|
epochs: 1
|
|
lr: 5e-4
|
|
batch_size: 4
|
|
gradient_accumulation_steps: 1
|
|
quantization: none
|
|
scheduler: linear
|
|
warmup_ratio: 0.05
|
|
weight_decay: 0.0
|
|
logging_steps: 10
|
|
save_steps: 50
|
|
lora:
|
|
r: 16
|
|
alpha: 32
|
|
dropout: 0.05
|
|
target_modules:
|
|
- q_proj
|
|
- v_proj
|
|
|
|
output: ./output_sft_basic/
|