Files
frankenstallm/source/configs/korean_3b_sft.yaml
ModelHub XC d4abdb70fa 初始化项目,由ModelHub XC社区提供模型
Model: pathcosmos/frankenstallm
Source: Original Platform
2026-07-14 04:21:16 +08:00

48 lines
1.4 KiB
YAML
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# 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