AWS SageMaker多轮RL微调搜索代理
Fine-tuning teaches a small search agent your tools and environment, giving it the reliability of a frontier model at lower latency and cost. In this post, we fine-tune an LLM-powe
All guides · Full article is primarily Simplified Chinese; use the English summary below for quick takeaways (GEO-friendly).
AWS SageMaker多轮RL微调搜索代理
Fine-tuning teaches a small search agent your tools and environment, giving it the reliability of a frontier model at lower latency and cost. In this post, we fine-tune an LLM-powered search agent using Amazon SageMaker AI multi-turn reinforcement learning (MTRL).
Who applies this? Enterprises building autonomous search agents that handle multi-hop queries across tools like BM25 lexical search and vector similarity, or any team needing reliable, cost-effective agentic behavior without frontier-model expense.
How to decide: Choose SageMaker MTRL when single-turn RL or supervised fine-tuning fails on interdependent multi-turn decisions; the trajectory-level reward captures full search sequences for better nDCG@10 and failure-rate reduction.
什么是AWS SageMaker AI多轮RL微调
Amazon SageMaker AI MTRL 是服务器端的多轮强化学习服务,可直接对LLM进行代理任务微调。
它将代理任务建模为决策序列,通过多轮rollout生成训练数据,并使用策略梯度算法优化模型。
核心优势包括:模块化代理-环境接口、低代码工具调用、异步rollout支持、原生PPO/CISPO算法库、恢复训练、轨迹可观测性以及评估作业。
这对搜索代理特别适用——奖励信号来自检索质量,多轮交互循环定义清晰,环境为搜索工具集。
准备环境与数据集
在AWS账户(us-west-2区域)中准备以下资源:
- 启用Amazon SageMaker AI,创建带BM25和向量搜索工具的代理端点。
- 将训练/验证数据集上传至S3,格式符合MTRL要求(包含prompt、response和完整轨迹)。
- 参考官方数据集:FRAMES、BRIGHT、Enterprise RAG、ESCI、Musique、MLQA用于训练,FreshStack、WixQA、BrowseComp-Plus、Wands用于测试。
训练配置与运行
启动MultiTurnRLTrainer:
from sagemaker.modules.train import MultiTurnRLTrainer
trainer = MultiTurnRLTrainer(
model_id="qwen3.6-27b",
agent_endpoint="your-agent-endpoint",
training_dataset_s3_uri="s3://your-bucket/datasets/train/",
validation_dataset_s3_uri="s3://your-bucket/datasets/validation/",
hyperparameters={
"max_epochs": 1,
"global_batch_size": 128,
"rollout_max_concurrency": 32,
},
output_s3_uri="s3://your-bucket/output/",
)
trainer.train
主要调整的参数:max_epochs(训练轮次)、global_batch_size(批次大小)、rollout_max_concurrency(并行rollout数)。默认算法与优势估计器已足够。
效果与验证
在BrowseComp-Plus基准上nDCG@10提升23.7%,WixQA提升18.4%,BrowseComp-Plus失败率从22.89%降至0.68%。训练曲线显示奖励持续上升并收敛,验证集表现稳定。
完整运行曲线与基准数据见AWS官方博客。
延伸阅读
- [AWS SageMaker MTRL官方文档]
- [多轮RL资产创建指南]
- [RooHome AI 价格雷达]
风险与边界
使用多轮RL微调搜索代理时,需注意资源成本(按Token计费)、数据隐私合规(AWS Region与合规标准匹配)及模型局限(非100%失败率消除)。这非法律意见,仅供参考。请以官方文档及AWS账户实际数据为准。