48 lines
1.4 KiB
YAML
48 lines
1.4 KiB
YAML
|
# Korean 3B SFT Configuration
|
|||
|
|
#
|
|||
|
|
# Base model: checkpoints/korean_3b_fp8_run1/checkpoint-XXXXXX (3B params pretrained)
|
|||
|
|
# SFT 목표: instruction following + 반복 퇴화 완화 + 생성 품질 향상
|
|||
|
|
# 아키텍처: LLaMA-3 3B 참고 (d=3072, 28L, 24H, GQA 8:1)
|
|||
|
|
#
|
|||
|
|
# 실행: bash scripts/launch_3b_sft.sh
|
|||
|
|
#
|
|||
|
|
# [설계 근거 — 2026-03-02]
|
|||
|
|
# - batch: 2 × 8GPU × 4 grad_accum = 64 eff_batch
|
|||
|
|
# - max_steps 33000 ≈ 3 epochs × 700K samples / 64 eff_batch
|
|||
|
|
# - lr=1e-5: pretrain 1.5e-4의 1/15 (catastrophic forgetting 방지)
|
|||
|
|
# - NEFTune alpha=5.0: 생성 다양성 향상, 반복 퇴화 완화
|
|||
|
|
# - use_fp8=true: B200 MXFP8 네이티브 가속 유지
|
|||
|
|
|
|||
|
|
model:
|
|||
|
|
vocab_size: 64000
|
|||
|
|
d_model: 3072
|
|||
|
|
n_layers: 28
|
|||
|
|
n_heads: 24
|
|||
|
|
n_kv_heads: 8
|
|||
|
|
d_ffn: 8192
|
|||
|
|
max_seq_len: 4096
|
|||
|
|
rope_theta: 500000.0
|
|||
|
|
dropout: 0.0
|
|||
|
|
bias: false
|
|||
|
|
use_flash_attn: true
|
|||
|
|
use_fp8: true
|
|||
|
|
|
|||
|
|
train:
|
|||
|
|
max_steps: 33000 # 3 epochs × 700K / 64 eff_batch
|
|||
|
|
batch_size: 2 # per GPU (3B VRAM 절약)
|
|||
|
|
grad_accum_steps: 4 # eff_batch: 2 × 8GPU × 4 = 64
|
|||
|
|
lr: 1.0e-5 # catastrophic forgetting 방지
|
|||
|
|
weight_decay: 0.01
|
|||
|
|
warmup_steps: 500
|
|||
|
|
max_grad_norm: 1.0
|
|||
|
|
log_interval: 10
|
|||
|
|
save_interval: 2000
|
|||
|
|
eval_interval: 500
|
|||
|
|
use_amp: false
|
|||
|
|
compile_model: false
|
|||
|
|
neftune_alpha: 5.0 # NEFTune noise injection
|
|||
|
|
|
|||
|
|
tokenizer:
|
|||
|
|
vocab_size: 64000
|
|||
|
|
type: sentencepiece_unigram
|