仅仅训练最后一轮对话

Former-commit-id: 1e7b396ff2

src/llamafactory/webui/components/train.py

@@ -44,10 +44,11 @@ def create_train_tab(engine: "Engine") -> Dict[str, "Component"]:
         )
         dataset_dir = gr.Textbox(value=DEFAULT_DATA_DIR, scale=1)
         dataset = gr.Dropdown(multiselect=True, allow_custom_value=True, scale=4)
+        train_last_turn_only = gr.Checkbox()
         preview_elems = create_preview_box(dataset_dir, dataset)
 
-    input_elems.update({training_stage, dataset_dir, dataset})
-    elem_dict.update(dict(training_stage=training_stage, dataset_dir=dataset_dir, dataset=dataset, **preview_elems))
+    input_elems.update({training_stage, dataset_dir, dataset,train_last_turn_only})
+    elem_dict.update(dict(training_stage=training_stage, dataset_dir=dataset_dir, dataset=dataset,train_last_turn_only=train_last_turn_only, **preview_elems))
 
     with gr.Row():
         learning_rate = gr.Textbox(value="5e-5")

src/llamafactory/webui/locales.py

@@ -536,6 +536,20 @@ LOCALES = {
             "info": "更改分词器词表和嵌入层的大小。",
         },
     },
+    "train_last_turn_only": {
+        "en": {
+            "label": "Train last turn only",
+            "info": "Train the model with the last turn only in multi turn.",
+        },
+        "ru": {
+            "label": "Обучать только последний поворот",
+            "info": "Обучать модель только последним поворотом в многоповоротном диалоге.",
+        },
+        "zh": {
+            "label": "仅最后一轮参与训练",
+            "info": "多轮对话仅使用最后一轮计算loss。",
+        },
+    },
     "use_llama_pro": {
         "en": {
             "label": "Enable LLaMA Pro",

src/llamafactory/webui/runner.py

@@ -125,6 +125,7 @@ class Runner:
             visual_inputs=get("top.visual_inputs"),
             dataset_dir=get("train.dataset_dir"),
             dataset=",".join(get("train.dataset")),
+            train_last_turn_only=get("train.train_last_turn_only"),
             cutoff_len=get("train.cutoff_len"),
             learning_rate=float(get("train.learning_rate")),
             num_train_epochs=float(get("train.num_train_epochs")),