-
Notifications
You must be signed in to change notification settings - Fork 0
Configuration
当你要把 DomainPostTrain 改成自己的领域、调整阶段输出或修改训练行为时，使用本页。
完整参数参考在 configs/README.md。本页只覆盖最常改的配置。
默认配置：
configs/domain_post_training.yaml
推荐工作流：
cp configs/domain_post_training.yaml configs/my_domain.yaml然后运行：
python scripts/training/train_pipeline.py --config configs/my_domain.yaml- 训练产物通常相对仓库根目录解析，例如
outputs/lora_adapter。 - 数据输入路径通常优先相对配置文件解析。默认配置中的
../data/...指向仓库级data/...。 -
null表示让代码使用默认行为。
| 配置项 | 为什么重要 |
|---|---|
base_model_repo_id |
download_models.py 使用的 Hugging Face Hub 仓库。 |
base_model_name_or_path |
训练、合并和推理实际加载的本地路径或模型 ID。 |
corpus.input_paths |
CPT 领域源文档。 |
fact_sft.input_paths |
Fact-SFT JSONL 数据。 |
dpo.input_path |
DPO 偏好数据，启用 DPO 时使用。 |
grpo.input_path |
GRPO 奖励提示数据，启用 GRPO 时使用。 |
eval.question_file |
训练后质量评估问题。 |
fact_sft.system_prompt |
领域角色、知识边界和安全边界。 |
fact_sft:
enabled: true
dpo:
enabled: false
grpo:
enabled: false完整流水线先运行 CPT，再运行已启用的后续阶段。也可以通过命令行跳过阶段，见 训练流水线。
显存紧张时，优先降低序列长度、LoRA rank，启用 4-bit 加载和 gradient checkpointing：
training:
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
max_seq_length: 768
load_in_4bit: true
gradient_checkpointing: true
peft:
r: 8
lora_alpha: 8训练验证集和训练后质量评估不是一回事：
- 验证集：训练过程中用于 loss/eval 信号。
- 质量评估：训练后检查事实回答、安全拒答和基础能力回归。
mock 数据很小，所以默认关闭验证：
corpus:
validation_mode: "none"
fact_sft:
validation_ratio: 0
dpo:
validation_ratio: 0真实项目建议使用独立 held-out CPT 文档：
corpus:
validation_mode: "separate_sources"
validation_sources:
- "../data/cpt_validation/source_documents"Use this page when adapting DomainPostTrain to another domain, changing stage outputs, or tuning training behavior.
The full parameter reference lives in configs/README.md. This page focuses on the settings users usually edit first.
The default config is:
configs/domain_post_training.yaml
Recommended workflow:
cp configs/domain_post_training.yaml configs/my_domain.yamlThen run commands with:
python scripts/training/train_pipeline.py --config configs/my_domain.yaml- Training artifacts usually resolve relative to the repository root, such as
outputs/lora_adapter. - Data input paths are usually resolved relative to the config file first. In the default config,
../data/...points back to repository-leveldata/.... -
nullmeans the code should use its default behavior.
| Setting | Why it matters |
|---|---|
base_model_repo_id |
Hugging Face Hub repo used by download_models.py. |
base_model_name_or_path |
Actual local path or model id loaded for training, merge, and inference. |
corpus.input_paths |
CPT source documents for the domain. |
fact_sft.input_paths |
Fact-SFT JSONL data. |
dpo.input_path |
DPO preference data, if DPO is enabled. |
grpo.input_path |
GRPO reward-prompt data, if GRPO is enabled. |
eval.question_file |
Post-training quality evaluation questions. |
fact_sft.system_prompt |
Domain role, knowledge boundary, and safety boundary. |
fact_sft:
enabled: true
dpo:
enabled: false
grpo:
enabled: falseThe full pipeline starts with CPT, then runs enabled later stages. You can also skip stages from the command line. See Training Pipeline.
When GPU memory is tight, start with smaller sequence length, smaller LoRA rank, 4-bit loading, and gradient checkpointing:
training:
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
max_seq_length: 768
load_in_4bit: true
gradient_checkpointing: true
peft:
r: 8
lora_alpha: 8Training validation and post-training quality evaluation are different:
- Validation set: used during training for loss/eval signal.
- Quality evaluation: run after training to check factual answers, safe refusals, and regressions.
The mock data is small, so validation is disabled by default:
corpus:
validation_mode: "none"
fact_sft:
validation_ratio: 0
dpo:
validation_ratio: 0For real projects, prefer independent held-out CPT documents:
corpus:
validation_mode: "separate_sources"
validation_sources:
- "../data/cpt_validation/source_documents"