模型训练教程与完整示例

所有语言版本中的命令和代码均保持英文。请将大写占位符替换为您账户和付费作业中的实际值；请勿将真实密钥粘贴到共享笔记本或日志中。

构建完整训练方案

实用示例应包含数据、计算资源和验收标准，而不只是提示词。

任务语言或业务目标
输入训练集、验证集和隐藏测试集
运行环境 兼容的计算资源与存储
结果验收标准与有效输出

训练可独立运行的德语模型

从配置章节中的完整自动模式配置开始。创建项目目录，放入兼容的学生模型、经过审核的训练材料、独立验证集和最终测试集，以及德语评估标准。选择服务商方案中包含的 reference model。SDK 应提出先决条件和主题，建立基线，并在获准预算内改进模型。

下方 JSON 用英文展示数据格式；请将实际审核过的德语文本放入真实数据文件，每行一个对象。
可运行的项目文件必须使用完整自动模式配置，不能只用方案片段。
分别衡量理解、写作、翻译和推理能力。不能保证小型学生模型在每项任务上都达到教师模型的水平。

{"text":"Customer: My invoice total is incorrect. Category: billing"}
{"text":"Customer: I cannot sign in. Category: account_access"}
{"text":"Customer: The application closes when I upload a file. Category: technical_issue"}

# Full Automatic German-language workflow.
# 1. Put a compatible local model in german-project/models/student/.
# 2. Save the complete project configuration as german-project/training.json.
# 3. Add independent train/validation/test files and the reviewed german rubric.
# 4. Replace provider/reference placeholders using your included model catalogue.
fpai-training doctor --config ./german-project/training.json
fpai-training plan --config ./german-project/training.json --dry-run
# The plan must break the objective into prerequisites and measurable topics.
fpai-training run --config ./german-project/training.json
fpai-training evaluate --run ./german-project/runs/german-001 --split test
fpai-training export --run ./german-project/runs/german-001 --format safetensors
# Evaluate comprehension, writing, translation and reasoning separately.
# A passed classifier does not prove all German-language capabilities.

将项目调整为业务任务

以下 JSON 示例是配置片段，不是完整项目文件。复制完整的自动或高级配置，用片段替换指定部分；保留身份信息、模型、教师模型、运行环境和有限预算，然后验证完整文件。分类和翻译需要不同的主题、数据与验收标准。

对于客服分类，应报告各类别的 precision 和 recall、macro F1、类别分布以及易混淆案例。
对于翻译，应使用独立的双语示例检查含义、术语、人名、数字和语气。
当 teacher_sharing 为 none 时，客户记录会留在本地。如果不允许共享的数据不足以实现目标，应停止并说明原因，而不是上传数据。

{
  "schema_version": "1.0", "mode": "automatic",
  "objective": "Classify support messages into billing, account_access or technical_issue.",
  "model": {"path": "./models/student", "adapter": "transformers-causal-lm", "revision": null},
  "teacher": {"provider": "openai", "reference_model_id": 123},
  "data": {"sources": ["./data/train.jsonl"], "teacher_sharing": "none", "validation_path": "./data/validation.jsonl", "test_path": "./data/test.jsonl", "deduplicate": true, "seed": 42},
  "runtime": {"target": "customer_datacenter", "devices": "auto", "distributed": {"backend": "none", "nodes": 1}},
  "training": {"method": "auto"},
  "curriculum": {"topics": ["intent", "ambiguous_requests", "class_balance"], "max_improvement_cycles": 3, "stop_on_regression": true},
  "budgets": {"max_training_minutes": 120, "max_teacher_requests": 100, "max_storage_gib": 20},
  "checkpoints": {"every_steps": 100, "keep_last": 3, "resume_validation": "strict"},
  "evaluation": {"rubric": "./rubric.json", "baseline_required": true},
  "output": {"directory": "./runs/support-001", "format": "adapters"}
}

{
  "objective": "Translate English support replies into German while preserving names, numbers and meaning.",
  "curriculum": {"topics": ["terminology", "meaning", "names_and_numbers", "tone"], "max_improvement_cycles": 3, "stop_on_regression": true},
  "evaluation": {"rubric": "./rubrics/translation.json", "baseline_required": true},
  "data": {"sources": ["./data/translation-train.jsonl"], "validation_path": "./data/translation-validation.jsonl", "test_path": "./data/translation-test.jsonl", "teacher_sharing": "approved_only"},
  "output": {"directory": "./runs/translation-001", "format": "safetensors"}
}

{"prompt":"My invoice total is wrong. Which category applies?","target":"billing","criterion":"support_category"}
{"prompt":"I cannot sign in. Which category applies?","target":"account_access","criterion":"support_category"}

{
  "schema_version": "1.0",
  "name": "support-category-evaluation",
  "scale": {"minimum": 0, "maximum": 1},
  "criteria": [
    {"id": "support_category", "description": "Match the expected category.", "metric": "macro_f1", "minimum_score": 0.85}
  ],
  "regression": {"maximum_score_drop": 0.02},
  "final_test": {"path": "./data/test.jsonl", "allow_training_use": false}
}

选择合适的基础设施并安全交付

这些运行环境示例也是完整审核配置中的片段。FusionPointAI Hosting、客户数据中心和兼容的 Colab 运行环境可提供计算资源；选择某个名称并不会分配 GPU、购买 notebook 点数，也不会使其他运行环境自动兼容。SDK 会在该运行环境中加载并训练学生模型。临时 notebook 结束前请保存输出。

托管示例：使用合同约定的计算环境、获批存储和账户权益；在适用情况下，托管属于单独的商业项目。
Colab 示例：将 API key 保存在 notebook 的机密信息存储中，检查可用 GPU 和存储，并在运行终止前将有效检查点复制到获批的持久存储。
分布式示例：使用获准的协调器和经过身份验证的节点通信，确保环境固定且 rendezvous/rank 值一致。launcher 只能管理本项目的 worker。

{
  "runtime": {"target": "fusionpointai_hosting", "devices": "auto", "distributed": {"backend": "none", "nodes": 1}},
  "budgets": {"max_training_minutes": 60, "max_teacher_requests": 500, "max_storage_gib": 20},
  "output": {"directory": "./runs/hosted-001", "format": "safetensors"}
}

{
  "runtime": {"target": "colab", "devices": [0], "distributed": {"backend": "none", "nodes": 1}},
  "checkpoints": {"every_steps": 50, "keep_last": 2, "resume_validation": "strict"},
  "budgets": {"max_training_minutes": 30, "max_teacher_requests": 100, "max_storage_gib": 10},
  "output": {"directory": "./runs/colab-001", "format": "adapters"}
}

{
  "runtime": {"target": "customer_datacenter", "devices": [0, 1], "distributed": {"backend": "ddp", "nodes": 2}},
  "checkpoints": {"every_steps": 100, "keep_last": 3, "resume_validation": "strict"},
  "output": {"directory": "./runs/distributed-001", "format": "safetensors"}
}

# Export a validated adapter or full model, not teacher weights.
fpai-training evaluate --run ./german-project/runs/support-001 --split test
fpai-training export --run ./german-project/runs/support-001 --format adapters
# Inspect artifact-manifest.json, pinned base/tokenizer and evaluation-report.json.
# Reload the package in a fresh compatible runtime before changing live inference.
# Export does not deploy, allocate hosting, or upload to an external model hub.

构建完整训练方案

训练可独立运行的德语模型

将项目调整为业务任务

选择合适的基础设施并安全交付

继续完成工作流程