Files
frankenstallm/source/configs/korean_1b_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

42 lines
1.2 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 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