模型合并测验

当您参加测验时，就会出现答案和解释。

1. 模型合并主要结合什么？
   • 多个模型的权重（参数）
   • 多个模型的训练数据集
   • 推理时多个模型的输出预测
   • 用于训练的硬件加速器
2. “任务向量”最好描述为：
   • 模型可以执行的任务列表
   • 微调模型的权重与其基本模型的权重之间的差异
   • 任务的输入标记向量
   • 在单个训练步骤中计算的梯度
3. 在描述的权重空间合并方法中，可以避免什么？
   • Evaluation of the merged checkpoint
   • Compatibility requirements between model weights
   • Additional gradient training and access to the original training data
   • Testing the result on intended tasks
4. TIES 合并和 DARE 在处理参数干扰方面有何不同？
   • Both use the same sign-voting method and differ only in scale
   • DARE uses prompt engineering while TIES retrains on joint data
   • Both average model outputs instead of changing weights
   • They use different techniques: TIES resolves sign conflicts, while DARE drops and rescales task-vector entries
5. 两个模型成功合并的关键前提是：
   • 共享相同的架构，并且通常源自公共基础模型
   • 使用相同的数据进行训练
   • 具有完全相同数量的训练步骤
   • 仅使用相同的分词器词汇量大小
6. 在任务算术中，系数在组合任务向量时控制什么？
   • Which tokenizer each checkpoint must use
   • How strongly each parameter-change vector contributes to the merged checkpoint
   • How many training examples are generated
   • Whether an evaluation set is required
7. TIES 合并命名的三个主要阶段是什么？
   • Tokenize text, infer emotions, and export audio
   • Quantize weights, shard GPUs, and route network packets
   • Trim small changes, elect a sign direction, then merge aligned parameter updates
   • Train a reward model, collect labels, and run PPO
8. Why should teams evaluate each task after a model merge?
   • Merging guarantees every benchmark improves
   • Evaluation is needed only to calculate the file size
   • A merged model cannot be run without retraining
   • Combining weight updates can create interference or reduce performance on a task