用swift微调Qwen3模型的function calling功能，如何构建数据集和模型参数以及如何避免过拟合和欠拟合

Checklist / 检查清单

• I have searched existing issues, and this is a new question or discussion topic. / 我已经搜索过现有的 issues，确认这是一个新的问题与讨论。

Question Description / 问题描述

这是数据集格式{"tools": [{"name": "query_ledger", "description": "查询电力台账数据。根据站点名称、母线名称、线路名称查询相关的电力设施台账信息。", "parameters": {"type": "object", "properties": {"subName": {"type": "string", "description": "站点名称"}}, "required": ["subName"]}}}], "messages": [{"role": "system", "content": "你是电网助手。先判断是否需要查询台账，再决定是否调用工具，最后给出结论。"}, {"role": "user", "content": "{"alert_content": "2025-08-14 20:19:23 某某站 退出(前置机A报警)", "alert_level": "事故", "signal_id": "", "signal_package": ["2025-08-14 20:19:23 某某站 退出(前置机A报警)"], "station_inside_outer": "1", "throbNum": "0"}"}, {"role": "assistant", "content": "\n某某"沙田站 退出(前置机A报警)"描述关键字"前置机A报警"判断告警类型为四遥中断,四遥中断类型的信号需要调用 query_ledger 查询站点台账进行辅助判断信号的正确性。\n从告警内容中提取站点名称为"某某站",调用 query_ledger 工具查询某某站的台账信息。\n", "tool_calls": [{"name": "query_ledger", "arguments": "{"subName": "沙田站", "busName": "", "lineName": ""}"}]}, {"role": "tool", "content": "[]"}, {"role": "assistant", "content": "经查询某某站不在台账中，说明沙田站不属于调管厂站，不需要关注某某站的任何信号，而沙田站 退出(前置机A报警)告警信号属于某某站不在调管范围，所以这条信号为正常信号。\n\n智能体查询到台账数量为0，说明台账中不存在相关某某站信息，说明某某站不属于调管厂站，不需要关注这个某某站告警信号，结果默认为正常信号。"}]}
这是模型参数/Users/e8900ai/anaconda3/envs/sft/bin/swift sft
--model /Users/e8900ai/LLM_Fine/models/qwen/Qwen3-4B
--tuner_type lora
--dataset /Users/e8900ai/LLM_Fine/models/data/data_update6.jsonl
--dataset /Users/e8900ai/LLM_Fine/models/data/general_agent_dataset.jsonl
--agent_template hermes
--split_dataset_ratio 0.1
--torch_dtype bfloat16
--num_train_epochs 2
--per_device_train_batch_size 4
--per_device_eval_batch_size 6
--learning_rate 1e-5
--gradient_accumulation_steps 4
--eval_steps 50
--save_steps 100
--save_total_limit 3
--logging_steps 10
--max_length 4096
--save_only_model true
--packing false
--attn_impl sdpa
--output_dir /Users/e8900ai/LLM_Fine/models/output4-4B
--lora_rank 8
--lora_alpha 16
--target_modules q_proj, k_proj, v_proj, o_proj
--lora_dropout 0.1
--lr_scheduler_type linear
--do_eval true
--load_best_model_at_end true
--eval_strategy steps
--metric_for_best_model loss
--train_sampling_strategy random
2>&1 | tee /Users/e8900ai/LLM_Fine/models/output4-4B/logs/training_$(date +%Y%m%d_%H%M%S).log

Review the dataset example and the supplied ms-swift sft command, including the LoRA, sampling, evaluation, and sequence-length settings. No repository file or test is named, so first locate the relevant fine-tuning and tool-calling documentation or entry points. Done means documenting validated dataset and parameter guidance that addresses overfitting and underfitting for this setup.