diff --git a/docs/en/notes/guide/mixer/tutorial.md b/docs/en/notes/guide/mixer/tutorial.md index f4d9b86..ab62edf 100644 --- a/docs/en/notes/guide/mixer/tutorial.md +++ b/docs/en/notes/guide/mixer/tutorial.md @@ -36,7 +36,7 @@ update_times: 2 - `init_mixture_proportions`: Initial sampling proportions, required when `mixture_sample_rule='mixture'`. - `warmup_step`: Before the first dynamic proportion update, the model needs to perform `warmup_step` steps of regular training. This helps the model establish initial understanding of data distribution. - `update_step`: Frequency of domain proportion updates. After every `update_step` training steps, the Mixer will be triggered to update domain proportions for the next training phase. -- `update_times`: Total number of dynamic data proportion calculations during the entire training process. Therefore, total training steps = `(update_times * update_step + warmup_step) * global_batch_size` +- `update_times`: Number of dynamic data proportion updates per Flex epoch. Total steps are derived from `num_train_epochs` unless `train_step > 0`. ### Static Mixing Configuration @@ -47,7 +47,7 @@ train_type: dynamic_mix static_mix: true # Whether to fix initial static mixing proportions (only effective in dynamic_mix trainer) mixture_sample_rule: mixture # Initial sampling rule init_mixture_proportions: [0.7, 0.3] # Initial proportions, can be adjusted by additional algorithms -train_step: 3 # Total training steps (only effective in dynamic_mix trainer), excluding warmup and update steps +train_step: 3 # fixed total steps; set to 0 to use num_train_epochs ``` When static mixing is enabled, the training process will use fixed `init_mixture_proportions` without dynamic adjustment. @@ -129,4 +129,4 @@ mixers: #### Key Points -- `params`: All parameters defined under this block will be passed as keyword arguments to the `__init__` constructor of the `RandomMixer` class. For example, the `seed` value here will be passed to the `seed` parameter of the `__init__` method. \ No newline at end of file +- `params`: All parameters defined under this block will be passed as keyword arguments to the `__init__` constructor of the `RandomMixer` class. For example, the `seed` value here will be passed to the `seed` parameter of the `__init__` method. diff --git a/docs/en/notes/guide/selector/quickstart.md b/docs/en/notes/guide/selector/quickstart.md index d90f7d1..d4cb1ce 100644 --- a/docs/en/notes/guide/selector/quickstart.md +++ b/docs/en/notes/guide/selector/quickstart.md @@ -23,6 +23,10 @@ component_name: less warmup_step: 4 update_step: 3 update_times: 2 +num_train_epochs: 1.0 +train_step: 0 eval_dataset: alpaca_zh_demo -``` \ No newline at end of file +``` + +`update_times` is the number of dynamic selections per Flex epoch. Use `num_train_epochs: 1.0` for one Flex epoch; use `num_train_epochs: N` with `train_step: 0` for multi-epoch runs. If `train_step > 0`, it fixes the total number of steps and overrides `num_train_epochs`. diff --git a/docs/en/notes/guide/selector/selector_delta_loss.md b/docs/en/notes/guide/selector/selector_delta_loss.md index e74d3d7..dda8a91 100644 --- a/docs/en/notes/guide/selector/selector_delta_loss.md +++ b/docs/en/notes/guide/selector/selector_delta_loss.md @@ -126,6 +126,8 @@ update_times: 2 eval_dataset: alpaca_zh_demo ``` +`update_times` is the number of dynamic selections per Flex epoch. Delta Loss requires `update_times >= 2`: the first selection builds initial losses, and later selections use delta loss. For multi-epoch runs, set `num_train_epochs: N` and keep `train_step: 0`. + --- ### Step 4: Run Training @@ -180,4 +182,3 @@ The merged model will be saved in: ## 3. Model Evaluation It is recommended to use the [DataFlow](https://github.com/OpenDCAI/DataFlow) [Model QA Evaluation Pipeline](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/) for systematic evaluation of the generated model. - diff --git a/docs/en/notes/guide/selector/selector_less.md b/docs/en/notes/guide/selector/selector_less.md index 57074c7..91b81f8 100644 --- a/docs/en/notes/guide/selector/selector_less.md +++ b/docs/en/notes/guide/selector/selector_less.md @@ -151,10 +151,10 @@ eval_dataset: alpaca_zh_demo * `output_dir`: Output directory of dynamic fine-tuning (LoRA adapter). * `warmup_step`: Number of warmup steps before the first sample selection. * `update_step`: Number of steps between each dynamic data selection. -* `update_times`: Total number of dynamic data selection iterations. +* `update_times`: Number of dynamic data selection iterations per Flex epoch. * `eval_dataset`: Validation dataset. -Both `dataset` and `eval_dataset` can be selected from `DataFlex/data/dataset_info.json` or local JSON files in ShareGPT/Alpaca format. Note: The training set size significantly affects computation cost. Total steps = `warmup_step + update_step × update_times`. +Both `dataset` and `eval_dataset` can be selected from `DataFlex/data/dataset_info.json` or local JSON files in ShareGPT/Alpaca format. Note: the training set size significantly affects computation cost. Steps per Flex epoch = `warmup_step + update_step × update_times`. Total steps are derived from `num_train_epochs` unless `train_step > 0`. --- @@ -212,4 +212,3 @@ The merged model will be saved in: ## 3. Model Evaluation It is recommended to use the [DataFlow](https://github.com/OpenDCAI/DataFlow) [Model QA Evaluation Pipeline](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/) for systematic evaluation of the generated model. - diff --git a/docs/en/notes/guide/selector/selector_loss.md b/docs/en/notes/guide/selector/selector_loss.md index 661cc83..f973aff 100644 --- a/docs/en/notes/guide/selector/selector_loss.md +++ b/docs/en/notes/guide/selector/selector_loss.md @@ -132,6 +132,8 @@ update_times: 2 eval_dataset: alpaca_zh_demo ``` +`update_times` is the number of dynamic selections per Flex epoch. For multi-epoch runs, set `num_train_epochs: N` and keep `train_step: 0`; if `train_step > 0`, it fixes total steps and overrides `num_train_epochs`. + --- ### Step 4: Run Training @@ -185,4 +187,3 @@ The merged model will be saved in: ## 3. Model Evaluation It is recommended to use the [DataFlow](https://github.com/OpenDCAI/DataFlow) [Model QA Evaluation Pipeline](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/) for systematic evaluation of the generated model. - diff --git a/docs/en/notes/guide/selector/selector_nice.md b/docs/en/notes/guide/selector/selector_nice.md index 5c77b0f..eeaa554 100644 --- a/docs/en/notes/guide/selector/selector_nice.md +++ b/docs/en/notes/guide/selector/selector_nice.md @@ -168,7 +168,7 @@ early_stopping_min_delta: 0.01 **Parameter description:** * `component_name`: Must match the `nice` component in `components.yaml`, determining reward backend and projection dimensions. -* `warmup_step` / `update_step` / `update_times`: Control the dynamic selection schedule; total steps = `warmup_step + update_step × update_times`. +* `warmup_step` / `update_step` / `update_times`: Control the dynamic selection schedule per Flex epoch. Total steps are derived from `num_train_epochs` unless `train_step > 0`. * `eval_dataset`: Validation set (Alpaca/ShareGPT style); reward model is used for scoring during generation. * `output_dir`: Path to save LoRA adapters and caches. @@ -226,4 +226,3 @@ The merged model will be saved to: It is recommended to use the [DataFlow](https://github.com/OpenDCAI/DataFlow) [Model QA Evaluation Pipeline](https://opendcai.github.io/DataFlow-Doc/en/guide/2k5wjgls/) to systematically evaluate the generated model, and to inspect the scoring logs in `cache_dir` to analyze the reward model’s sensitivity to different samples. - diff --git a/docs/en/notes/guide/selector/selector_offline_near.md b/docs/en/notes/guide/selector/selector_offline_near.md index 2e967c7..3c5936f 100644 --- a/docs/en/notes/guide/selector/selector_offline_near.md +++ b/docs/en/notes/guide/selector/selector_offline_near.md @@ -153,7 +153,7 @@ update_times: 2 **Notes:** * `component_name: near` enables the NEAR component. -* `warmup_step / update_step / update_times` decide **when** and **how often** to re‑select the training subset; total steps ≈ `warmup_step + update_step × update_times`. +* `warmup_step / update_step / update_times` decide **when** and **how often** to re-select the training subset in each Flex epoch. Total steps are derived from `num_train_epochs` unless `train_step > 0`. * total batch_size=device_number x per_device_train_batch_size x gradient_accumulation_steps --- @@ -201,4 +201,3 @@ llamafactory-cli export llama3_lora_sft.yaml We recommend using the [DataFlow](https://github.com/OpenDCAI/DataFlow) QA evaluation pipeline to compare **NEAR** against **Less** and **random sampling**. - diff --git a/docs/en/notes/guide/selector/selector_offline_tsds.md b/docs/en/notes/guide/selector/selector_offline_tsds.md index 137ced9..9168b8f 100644 --- a/docs/en/notes/guide/selector/selector_offline_tsds.md +++ b/docs/en/notes/guide/selector/selector_offline_tsds.md @@ -187,7 +187,7 @@ update_times: 2 **Notes:** * `component_name: tsds` enables the TSDS component. -* `warmup_step / update_step / update_times` decide **when** and **how often** to re‑select the training subset; total steps ≈ `warmup_step + update_step × update_times`. +* `warmup_step / update_step / update_times` decide **when** and **how often** to re-select the training subset in each Flex epoch. Total steps are derived from `num_train_epochs` unless `train_step > 0`. * total batch_size=device_number x per_device_train_batch_size x gradient_accumulation_steps --- @@ -235,4 +235,3 @@ llamafactory-cli export llama3_lora_sft.yaml We recommend using the [DataFlow](https://github.com/OpenDCAI/DataFlow) QA evaluation pipeline to compare **TSDS** against **Less** and **random sampling**. - diff --git a/docs/en/notes/guide/selector/selector_zeroth.md b/docs/en/notes/guide/selector/selector_zeroth.md index 5d588d3..83247d9 100644 --- a/docs/en/notes/guide/selector/selector_zeroth.md +++ b/docs/en/notes/guide/selector/selector_zeroth.md @@ -123,10 +123,10 @@ eval_dataset: alpaca_zh_demo * `output_dir`: Output directory of dynamic fine-tuning (LoRA adapter). * `warmup_step`: Number of warmup steps before the first sample selection. * `update_step`: Number of steps between each dynamic data selection. -* `update_times`: Total number of dynamic data selection iterations. +* `update_times`: Number of dynamic data selection iterations per Flex epoch. * `eval_dataset`: Validation dataset. -Both `dataset` and `eval_dataset` can be selected from `DataFlex/data/dataset_info.json` or local JSON files in ShareGPT/Alpaca format. Note: The training set size significantly affects computation cost. Total steps = `warmup_step + update_step × update_times`. +Both `dataset` and `eval_dataset` can be selected from `DataFlex/data/dataset_info.json` or local JSON files in ShareGPT/Alpaca format. Note: the training set size significantly affects computation cost. Steps per Flex epoch = `warmup_step + update_step × update_times`. Total steps are derived from `num_train_epochs` unless `train_step > 0`. --- diff --git a/docs/en/notes/guide/weighter/quickstart.md b/docs/en/notes/guide/weighter/quickstart.md index 69fa692..ba8db81 100644 --- a/docs/en/notes/guide/weighter/quickstart.md +++ b/docs/en/notes/guide/weighter/quickstart.md @@ -21,5 +21,8 @@ train_type: dynamic_weight components_cfg_file: src/dataflex/configs/components.yaml component_name: loss warmup_step: 1 -train_step: 3 # total train steps (including warmup) -``` \ No newline at end of file +num_train_epochs: 1.0 +train_step: 0 # set positive to fix total steps and override num_train_epochs +``` + +For multi-epoch runs, set `num_train_epochs: N` and keep `train_step: 0`. `warmup_step` is a global step threshold and does not reset each epoch. diff --git a/docs/en/notes/guide/weighter/tutorial.md b/docs/en/notes/guide/weighter/tutorial.md index de05afb..b6f92d8 100644 --- a/docs/en/notes/guide/weighter/tutorial.md +++ b/docs/en/notes/guide/weighter/tutorial.md @@ -26,7 +26,8 @@ train_type: dynamic_weight # Select trainer type. Available options: components_cfg_file: src/dataflex/configs/components.yaml component_name: loss # Select component name, corresponding to components defined in components_cfg_file warmup_step: 1 -train_step: 3 # Total training steps (including warm_up) +num_train_epochs: 1.0 +train_step: 0 # set positive to fix total steps and override num_train_epochs ``` ### Parameter Details @@ -35,7 +36,8 @@ train_step: 3 # Total training steps (including warm_up) - `component_name`: Defines the specific strategy for data weighting. For example, `loss` uses a loss-based weighter. - `components_cfg_file`: Defines the parameter file containing specific parameters for the corresponding strategy. - `warmup_step`: Before the first dynamic weighting, the model needs to perform `warmup_step` steps of regular training. This helps the model establish initial understanding of data distribution. -- `train_step`: Total training steps (including warmup). Weight Trainer will dynamically weight samples at each training step after warmup completion. +- `train_step`: Optional fixed total steps. If `train_step > 0`, it overrides `num_train_epochs`; for multi-epoch runs, keep `train_step: 0`. +- `num_train_epochs`: Controls the number of epochs when `train_step: 0`. `warmup_step` is a global step threshold and does not reset each epoch. ## How to Add Custom Weighter in DataFlex @@ -140,4 +142,4 @@ weighters: #### Key Points -- `params`: All parameters defined under this block will be passed as keyword arguments to the `__init__` constructor of the `CustomWeighter` class. For example, the `strategy` value here will be passed to the `strategy` parameter of the `__init__` method. \ No newline at end of file +- `params`: All parameters defined under this block will be passed as keyword arguments to the `__init__` constructor of the `CustomWeighter` class. For example, the `strategy` value here will be passed to the `strategy` parameter of the `__init__` method. diff --git a/docs/zh/notes/guide/mixer/tutorial.md b/docs/zh/notes/guide/mixer/tutorial.md index c87d9dd..f271d54 100644 --- a/docs/zh/notes/guide/mixer/tutorial.md +++ b/docs/zh/notes/guide/mixer/tutorial.md @@ -36,7 +36,7 @@ update_times: 2 - `init_mixture_proportions`: 初始采样对应的比例，`mixture_sample_rule='mixture'` 时需要指定。 - `warmup_step`: 在执行第一次动态配比更新前，模型需要先进行 `warmup_step` 步的常规训练。这有助于模型建立对数据分布的初步认知。 - `update_step`: 领域配比更新的频率。每当训练进行 `update_step` 步后，Mixer 将被触发，更新领域配比用于下一阶段的训练。 -- `update_times`: 整个训练过程中，动态数据配比计算的总次数。因此总的训练步数为 `(update_times * update_step + warmup_step) * global_batch_size` +- `update_times`: 每个 Flex epoch 内动态数据配比计算的次数。总步数由 `num_train_epochs` 推导；若 `train_step > 0`，则以 `train_step` 为准。 ### 静态混合配置 @@ -47,7 +47,7 @@ train_type: dynamic_mix static_mix: true # 是否固定初始静态混合比例（仅在dynamic_mix训练器中生效） mixture_sample_rule: mixture # 初始采样规则 init_mixture_proportions: [0.7, 0.3] # 对应初始的比例，可通过额外算法自行调整 -train_step: 3 # 总训练步数（仅在dynamic_mix训练器中生效），不考虑warmup和update steps +train_step: 3 # 固定总步数；设为 0 时由 num_train_epochs 控制 ``` 启用静态混合后，训练过程中将使用固定的 `init_mixture_proportions` 比例，不再动态调整。 @@ -129,4 +129,4 @@ mixers: #### 关键点说明 -- `params`: 该块下定义的所有参数都将作为关键字参数传递给 `RandomMixer` 类的 `__init__` 构造函数。例如，这里的 `seed` 值会传递给 `__init__` 方法的 `seed` 参数。 \ No newline at end of file +- `params`: 该块下定义的所有参数都将作为关键字参数传递给 `RandomMixer` 类的 `__init__` 构造函数。例如，这里的 `seed` 值会传递给 `__init__` 方法的 `seed` 参数。 diff --git a/docs/zh/notes/guide/selector/quickstart.md b/docs/zh/notes/guide/selector/quickstart.md index d426da1..30dfe60 100644 --- a/docs/zh/notes/guide/selector/quickstart.md +++ b/docs/zh/notes/guide/selector/quickstart.md @@ -23,6 +23,10 @@ component_name: less warmup_step: 4 update_step: 3 update_times: 2 +num_train_epochs: 1.0 +train_step: 0 eval_dataset: alpaca_zh_demo ``` + +`update_times` 表示每个 Flex epoch 内的动态选择次数。单个 Flex epoch 使用 `num_train_epochs: 1.0`；多 epoch 使用 `num_train_epochs: N` 且保持 `train_step: 0`。若 `train_step > 0`，则固定总步数并覆盖 `num_train_epochs`。 diff --git a/docs/zh/notes/guide/selector/selector_delta_loss.md b/docs/zh/notes/guide/selector/selector_delta_loss.md index d89333d..0eedd30 100644 --- a/docs/zh/notes/guide/selector/selector_delta_loss.md +++ b/docs/zh/notes/guide/selector/selector_delta_loss.md @@ -127,6 +127,8 @@ update_times: 2 eval_dataset: alpaca_zh_demo ``` +`update_times` 表示每个 Flex epoch 内的动态选择次数。Delta Loss 需要 `update_times >= 2`：第一次选择建立 initial loss，后续选择根据 delta loss 更新样本。多 epoch 训练请设置 `num_train_epochs: N` 且保持 `train_step: 0`。 + --- ### 步骤四：运行训练 @@ -178,4 +180,3 @@ llamafactory-cli export llama3_lora_sft.yaml ## 3. 模型评估 推荐使用[DataFlow](https://github.com/OpenDCAI/DataFlow)的[模型QA能力评估流水线](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/)对生成后的模型进行系统性评估。 - diff --git a/docs/zh/notes/guide/selector/selector_less.md b/docs/zh/notes/guide/selector/selector_less.md index 935ea85..8e7c1fe 100644 --- a/docs/zh/notes/guide/selector/selector_less.md +++ b/docs/zh/notes/guide/selector/selector_less.md @@ -146,12 +146,12 @@ eval_dataset: alpaca_zh_demo * `output_dir`: 动态微调结果（LoRA 适配器）的输出路径。 * `warmup_step`: 训练初期第一次训练数据选择前，进行warmup的步数。 * `update_step`: 每次训练数据动态选择的步数。 -* `update_times`: 数据动态选择的总次数。 +* `update_times`: 每个 Flex epoch 内的数据动态选择次数。 * `eval_dataset`: 验证数据集。 dataset和eval_dataset可选`DataFlex/data/dataset_info.json`中数据，或本地路径下sharegpt或alpaca格式的json数据。注意该方法的情形下，训练集规模会较大影响计算成本。 -总步数 = `warmup_step + update_step × update_times`。 +每个 Flex epoch 的步数 = `warmup_step + update_step × update_times`。总步数由 `num_train_epochs` 推导；若 `train_step > 0`，则以 `train_step` 为准。 --- @@ -207,4 +207,3 @@ llamafactory-cli export llama3_lora_sft.yaml ## 3. 模型评估 推荐使用[DataFlow](https://github.com/OpenDCAI/DataFlow)的[模型QA能力评估流水线](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/)对生成后的模型进行系统性评估。 - diff --git a/docs/zh/notes/guide/selector/selector_loss.md b/docs/zh/notes/guide/selector/selector_loss.md index 205a5a8..964b43e 100644 --- a/docs/zh/notes/guide/selector/selector_loss.md +++ b/docs/zh/notes/guide/selector/selector_loss.md @@ -132,6 +132,8 @@ update_times: 2 eval_dataset: alpaca_zh_demo ``` +`update_times` 表示每个 Flex epoch 内的动态选择次数。多 epoch 训练请设置 `num_train_epochs: N` 且保持 `train_step: 0`；若 `train_step > 0`，则固定总步数并覆盖 `num_train_epochs`。 + --- ### 步骤四：运行训练 @@ -184,4 +186,3 @@ llamafactory-cli export llama3_lora_sft.yaml ## 3. 模型评估 推荐使用[DataFlow](https://github.com/OpenDCAI/DataFlow)的[模型QA能力评估流水线](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/)对生成后的模型进行系统性评估。 - diff --git a/docs/zh/notes/guide/selector/selector_nice.md b/docs/zh/notes/guide/selector/selector_nice.md index 51c8ea0..9304e01 100644 --- a/docs/zh/notes/guide/selector/selector_nice.md +++ b/docs/zh/notes/guide/selector/selector_nice.md @@ -162,7 +162,7 @@ early_stopping_min_delta: 0.01 **参数说明：** * `component_name`: 与 `components.yaml` 中的 `nice` 组件保持一致，决定奖励后端与投影维度等设置。 -* `warmup_step` / `update_step` / `update_times`: 决定动态选择的触发节奏；总步数 = `warmup_step + update_step × update_times`。 +* `warmup_step` / `update_step` / `update_times`: 决定每个 Flex epoch 内的动态选择节奏；总步数由 `num_train_epochs` 推导，若 `train_step > 0` 则以 `train_step` 为准。 * `eval_dataset`: 验证集，可以是 Alpaca/ShareGPT 样式，生成时会调用奖励模型打分。 * `output_dir`: LoRA 适配器与缓存保存路径。 diff --git a/docs/zh/notes/guide/selector/selector_offline_near.md b/docs/zh/notes/guide/selector/selector_offline_near.md index 6de560e..44ce20b 100644 --- a/docs/zh/notes/guide/selector/selector_offline_near.md +++ b/docs/zh/notes/guide/selector/selector_offline_near.md @@ -173,7 +173,7 @@ update_times: 2 **参数说明：** * `component_name: near`：启用 NEAR 组件。 -* `warmup_step / update_step / update_times`：决定**何时**与**多久**进行一次动态选择；总步数 ≈ `warmup_step + update_step × update_times`。 +* `warmup_step / update_step / update_times`：决定每个 Flex epoch 内**何时**与**多久**进行一次动态选择；总步数由 `num_train_epochs` 推导，若 `train_step > 0` 则以 `train_step` 为准。 * 总batch_size=device_number x per_device_train_batch_size x gradient_accumulation_steps @@ -218,4 +218,4 @@ llamafactory-cli export llama3_lora_sft.yaml ## 9. 评估与对比 -建议使用 [DataFlow](https://github.com/OpenDCAI/DataFlow) 的模型 QA 评估流水线，对 **NEAR** 与 **Less**、**随机采样** 等策略进行并列评测 \ No newline at end of file +建议使用 [DataFlow](https://github.com/OpenDCAI/DataFlow) 的模型 QA 评估流水线，对 **NEAR** 与 **Less**、**随机采样** 等策略进行并列评测 diff --git a/docs/zh/notes/guide/selector/selector_offline_tsds.md b/docs/zh/notes/guide/selector/selector_offline_tsds.md index 9bdaa2e..82e606c 100644 --- a/docs/zh/notes/guide/selector/selector_offline_tsds.md +++ b/docs/zh/notes/guide/selector/selector_offline_tsds.md @@ -209,7 +209,7 @@ update_times: 2 **参数说明：** * `component_name: tsds`：启用 TSDS 组件。 -* `warmup_step / update_step / update_times`：决定**何时**与**多久**进行一次动态选择；总步数 ≈ `warmup_step + update_step × update_times`。 +* `warmup_step / update_step / update_times`：决定每个 Flex epoch 内**何时**与**多久**进行一次动态选择；总步数由 `num_train_epochs` 推导，若 `train_step > 0` 则以 `train_step` 为准。 * 总batch_size=device_number x per_device_train_batch_size x gradient_accumulation_steps --- @@ -253,4 +253,4 @@ llamafactory-cli export llama3_lora_sft.yaml ## 9. 评估与对比 -建议使用 [DataFlow](https://github.com/OpenDCAI/DataFlow) 的模型 QA 评估流水线，对 **TSDS** 与 **Less**、**随机采样** 等策略进行并列评测 \ No newline at end of file +建议使用 [DataFlow](https://github.com/OpenDCAI/DataFlow) 的模型 QA 评估流水线，对 **TSDS** 与 **Less**、**随机采样** 等策略进行并列评测 diff --git a/docs/zh/notes/guide/selector/selector_zeroth.md b/docs/zh/notes/guide/selector/selector_zeroth.md index ba59137..24c0995 100644 --- a/docs/zh/notes/guide/selector/selector_zeroth.md +++ b/docs/zh/notes/guide/selector/selector_zeroth.md @@ -132,12 +132,12 @@ eval_dataset: alpaca_zh_demo * `output_dir`: 动态微调结果（LoRA 适配器）的输出路径。 * `warmup_step`: 训练初期第一次训练数据选择前，进行warmup的步数。 * `update_step`: 每次训练数据动态选择的步数。 -* `update_times`: 数据动态选择的总次数。 +* `update_times`: 每个 Flex epoch 内的数据动态选择次数。 * `eval_dataset`: 验证数据集。 dataset和eval_dataset可选`DataFlex/data/dataset_info.json`中数据，或本地路径下sharegpt或alpaca格式的json数据。注意该方法的情形下，训练集规模会较大影响计算成本。 -总步数 = `warmup_step + update_step × update_times`。 +每个 Flex epoch 的步数 = `warmup_step + update_step × update_times`。总步数由 `num_train_epochs` 推导；若 `train_step > 0`，则以 `train_step` 为准。 --- @@ -156,4 +156,3 @@ FORCE_TORCHRUN=1 DISABLE_VERSION_CHECK=1 dataflex-cli train examples/train_lora/ ## 3. 模型评估 推荐使用[DataFlow](https://github.com/OpenDCAI/DataFlow)的[模型QA能力评估流水线](https://opendcai.github.io/DataFlow-Doc/zh/guide/2k5wjgls/)对生成后的模型进行系统性评估。 - diff --git a/docs/zh/notes/guide/weighter/quickstart.md b/docs/zh/notes/guide/weighter/quickstart.md index 8c09ef0..99456a4 100644 --- a/docs/zh/notes/guide/weighter/quickstart.md +++ b/docs/zh/notes/guide/weighter/quickstart.md @@ -21,5 +21,8 @@ train_type: dynamic_weight components_cfg_file: src/dataflex/configs/components.yaml component_name: loss warmup_step: 1 -train_step: 3 # 总训练步数（包括warm_up） -``` \ No newline at end of file +num_train_epochs: 1.0 +train_step: 0 # 设为正数时固定总步数，并覆盖 num_train_epochs +``` + +多 epoch 训练只需设置 `num_train_epochs: N` 并保持 `train_step: 0`。`warmup_step` 是全局 step 阈值，不会在每个 epoch 重置。 diff --git a/docs/zh/notes/guide/weighter/tutorial.md b/docs/zh/notes/guide/weighter/tutorial.md index e1e6658..8df0236 100644 --- a/docs/zh/notes/guide/weighter/tutorial.md +++ b/docs/zh/notes/guide/weighter/tutorial.md @@ -26,7 +26,8 @@ train_type: dynamic_weight # 选择训练器类型。可选值包括： components_cfg_file: src/dataflex/configs/components.yaml component_name: loss # 选择组件名称，对应 components_cfg_file 中定义的组件 warmup_step: 1 -train_step: 3 # 总训练步数（包括warm_up） +num_train_epochs: 1.0 +train_step: 0 # 设为正数时固定总步数，并覆盖 num_train_epochs ``` ### 参数详解 @@ -35,7 +36,8 @@ train_step: 3 # 总训练步数（包括warm_up） - `component_name`: 定义数据加权的具体策略。例如，`loss` 表示使用基于损失值的加权器。 - `components_cfg_file`: 定义策略的参数文件，包含对应策略的特定参数。 - `warmup_step`: 在执行第一次动态加权前，模型需要先进行 `warmup_step` 步的常规训练。这有助于模型建立对数据分布的初步认知。 -- `train_step`: 总训练步数（包括 warmup），Weight Trainer 将在 warmup 完成后的每个训练步骤中对样本进行动态加权。 +- `train_step`: 可选的固定总步数。`train_step > 0` 时覆盖 `num_train_epochs`；多 epoch 训练请保持 `train_step: 0`。 +- `num_train_epochs`: 当 `train_step: 0` 时控制训练 epoch 数。`warmup_step` 是全局 step 阈值，不会在每个 epoch 重置。 ## 如何在 DataFlex 中添加自定义 Weighter @@ -140,4 +142,4 @@ weighters: #### 关键点说明 -- `params`: 该块下定义的所有参数都将作为关键字参数传递给 `CustomWeighter` 类的 `__init__` 构造函数。例如，这里的 `strategy` 值会传递给 `__init__` 方法的 `strategy` 参数。 \ No newline at end of file +- `params`: 该块下定义的所有参数都将作为关键字参数传递给 `CustomWeighter` 类的 `__init__` 构造函数。例如，这里的 `strategy` 值会传递给 `__init__` 方法的 `strategy` 参数。