Guanghu Domestic Migration d1e47f4565
Some checks are pending
自动更新代码和重启 / update-and-restart (push) Waiting to run
CI检查 + 自动部署 / check (push) Waiting to run
CI检查 + 自动部署 / deploy (push) Blocked by required conditions
重启聊天服务 / restart (push) Waiting to run
chore: import sanitized domestic snapshot for REPO-002
Source snapshot: ca48d3ddf926d79aa138306164169baf764bb829
2026-07-17 15:54:41 +08:00

47 lines
1.7 KiB
YAML
Raw 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.

# 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