42 lines
1.2 KiB
YAML
42 lines
1.2 KiB
YAML
# Korean LLM 1B — SFT (Supervised Fine-Tuning) 설정
|
||
#
|
||
# Base model: korean_1b_fp8_run1/checkpoint-0034000 (1.19B params, 34k pretrain steps)
|
||
# SFT 목표: instruction following + 반복 퇴화 완화 + 생성 품질 향상
|
||
#
|
||
# 실행: bash scripts/launch_sft.sh
|
||
|
||
model:
|
||
vocab_size: 64000
|
||
d_model: 2048
|
||
n_layers: 24
|
||
n_heads: 16
|
||
n_kv_heads: 4
|
||
d_ffn: 5472
|
||
max_seq_len: 4096
|
||
rope_theta: 500000.0
|
||
dropout: 0.0
|
||
bias: false
|
||
use_flash_attn: true
|
||
use_fp8: true
|
||
|
||
train:
|
||
max_steps: 5000 # SFT: 수천 steps면 충분 (pretrain 34k 대비 ~10%)
|
||
batch_size: 4 # per GPU (SFT는 seq가 다양하므로 작게)
|
||
grad_accum_steps: 2 # eff_batch: 4 × 8GPU × 2 × 4096 = 262,144 tok/step
|
||
lr: 2.0e-5 # pretrain의 1/10 (catastrophic forgetting 방지)
|
||
weight_decay: 0.01 # pretrain 0.1보다 약하게
|
||
warmup_steps: 150 # 3000 steps의 3.3%
|
||
max_grad_norm: 1.0
|
||
log_interval: 10
|
||
save_interval: 500
|
||
eval_interval: 100
|
||
use_amp: false # FP8 사용 시 불필요
|
||
compile_model: false
|
||
fp8_amax_history_len: 16
|
||
fp8_amax_compute_algo: "max"
|
||
fp8_format: "MXFP8"
|
||
|
||
tokenizer:
|
||
vocab_size: 64000
|
||
type: sentencepiece_unigram
|