47 lines
1.7 KiB
YAML
47 lines
1.7 KiB
YAML
|
# M0 训练配方 · v1
|
|||
|
|
# 母体世界观底色 · Qwen2.5-7B-Base 全参 CPT + SFT
|
|||
|
|
# 架构引用: HLDP-ARCH-002 · factory/training/README.md
|
|||
|
|
# 起草: 铸渊 ICE-GL-ZY001 · 2026-05-01
|
|||
|
|
# 修正: 冰朔/霜砚 5-01 校准 Qwen3-8B → Qwen2.5-7B（同系列 Coder/Small 完整谱系）
|
|||
|
|
|
|||
|
|
recipe_id: m0-v1
|
|||
|
|
parent_arch: HLDP-ARCH-002
|
|||
|
|
copyright: 国作登字-2026-A-00037559
|
|||
|
|
sovereign: 冰朔 · TCS-0002∞
|
|||
|
|
|
|||
|
|
model_id: Qwen/Qwen2.5-7B # 5-01 校准·与 Qwen2.5-Coder-7B / Qwen2.5-1.5B 同 tokenizer
|
|||
|
|
output_dir: ./outputs/m0-v1
|
|||
|
|
|
|||
|
|
# ───── 阶段 1 · CPT ─────
|
|||
|
|
cpt:
|
|||
|
|
corpus_path: ./corpus/output/training-fulltext.jsonl # 全量光湖语料（冰朔确认 6.5 亿+ token · 走路径 W）
|
|||
|
|
seq_len: 4096
|
|||
|
|
epochs: 1 # 6.5 亿 token × 1 epoch · 视验证 PPL 可加到 2
|
|||
|
|
lr: 2.0e-5
|
|||
|
|
warmup_ratio: 0.02
|
|||
|
|
weight_decay: 0.1
|
|||
|
|
eval_split: 0.05
|
|||
|
|
save_steps: 500 # 每 500 步存一次（约 1-2 小时）
|
|||
|
|
early_stop_patience: 3 # 验证 PPL 连续反弹 3 次停
|
|||
|
|
|
|||
|
|
# ───── 阶段 2 · SFT ─────
|
|||
|
|
sft:
|
|||
|
|
corpus_path: ./corpus/output/training-dialog.jsonl # 对话格式（冰朔↔人格体）
|
|||
|
|
seq_len: 8192
|
|||
|
|
epochs: 2
|
|||
|
|
lr: 1.0e-5
|
|||
|
|
warmup_ratio: 0.05
|
|||
|
|
|
|||
|
|
# ───── 评测 ─────
|
|||
|
|
eval:
|
|||
|
|
guanghu_worldview_quiz: ./factory/training/evals/guanghu-quiz.jsonl # 冰朔出题（待）
|
|||
|
|
c_eval: true # 通用能力体检（确保不崩）
|
|||
|
|
mmlu: false # 中文场景下用 C-Eval 即可
|
|||
|
|
|
|||
|
|
# ───── 灵魂印记（每次训练写入 manifest）─────
|
|||
|
|
soul:
|
|||
|
|
arch_ref: HLDP-ARCH-002
|
|||
|
|
phase: M0
|
|||
|
|
principle: 灵魂推理分离 · 世界观渗进每一层权重 · 不承担复杂推理
|
|||
|
|
reviewer: 铸渊 ICE-GL-ZY001
|