update

app.py

@@ -20,11 +20,11 @@ DEFAULT_FUSER_MAX_NEW_TOKENS = 256
 
 
 # MIX-INSTRUCT
-EXAMPLES_DATASET = load_dataset("llm-blender/mix-instruct", split='validation')
+EXAMPLES_DATASET = load_dataset("llm-blender/mix-instruct", split='validation', streaming=True)
 SHUFFLED_EXAMPLES_DATASET = EXAMPLES_DATASET.shuffle(seed=42)
 MIX_INSTRUCT_EXAMPLES = []
 CANDIDATE_MAP = {}
-for i, example in enumerate(SHUFFLED_EXAMPLES_DATASET):
+for i, example in enumerate(SHUFFLED_EXAMPLES_DATASET.take(100)):
     MIX_INSTRUCT_EXAMPLES.append([
         example['instruction'],
         example['input'],
@@ -88,7 +88,7 @@ def get_mt_bench_human_judge_examples(model_a, model_b, conversation_a, conversa
 
 blender = llm_blender.Blender()
 blender.loadranker("llm-blender/PairRM")
-blender.loadfuser("llm-blender/gen_fuser_3b")
+# blender.loadfuser("llm-blender/gen_fuser_3b")
 
 def update_base_llms_num(k, llm_outputs):
     k = int(k)
@@ -214,9 +214,9 @@ with gr.Blocks(theme='ParityError/Anime') as demo:
                 outputs=[compare_result, compare_result_prob],
             )
             clear_button.click(
-                fn=lambda: ["", ""],
+                fn=lambda: ["", "", "", "", ""],
                 inputs=[],
-                outputs=[compare_result, compare_result_prob],
+                outputs=[instruction, response1, response2, compare_result, compare_result_prob],
             )
             
             hhh_dummy_textbox1 = gr.Textbox(lines=1, label="subset", placeholder="", show_label=False, visible=False)
@@ -328,207 +328,207 @@ with gr.Blocks(theme='ParityError/Anime') as demo:
             )
     
     
-    with gr.Tab("LLM-Blender"):
-        # llm-blender interface 
-        with gr.Row():
-            gr.Markdown(descriptions.LLM_BLENDER_OVERALL_DESC)
-            gr.Image("https://github.com/yuchenlin/LLM-Blender/blob/main/docs/llm_blender.png?raw=true", height=300)
-        gr.Markdown("## Input and Base LLMs")
-        with gr.Row():
-            with gr.Column():
-                inst_textbox = gr.Textbox(lines=1, label="Instruction", placeholder="Enter instruction here", show_label=True)
-                input_textbox = gr.Textbox(lines=4, label="Input Context", placeholder="Enter input context here", show_label=True)
-            with gr.Column():
-                saved_llm_outputs = gr.State(value={})
-                with gr.Group():
-                    selected_base_llm_name_dropdown = gr.Dropdown(label="Base LLM",
-                        choices=[f"LLM-{i+1}" for i in range(MIN_BASE_LLM_NUM)], value="LLM-1", show_label=True)
-                    selected_base_llm_output = gr.Textbox(lines=4, label="LLM-1 (Click Save to save current content)",
-                        placeholder="Enter LLM-1 output here", show_label=True)
-                with gr.Row():
-                    base_llm_outputs_save_button = gr.Button('Save', variant='primary')
+    # with gr.Tab("LLM-Blender"):
+    #     # llm-blender interface 
+    #     with gr.Row():
+    #         gr.Markdown(descriptions.LLM_BLENDER_OVERALL_DESC)
+    #         gr.Image("https://github.com/yuchenlin/LLM-Blender/blob/main/docs/llm_blender.png?raw=true", height=300)
+    #     gr.Markdown("## Input and Base LLMs")
+    #     with gr.Row():
+    #         with gr.Column():
+    #             inst_textbox = gr.Textbox(lines=1, label="Instruction", placeholder="Enter instruction here", show_label=True)
+    #             input_textbox = gr.Textbox(lines=4, label="Input Context", placeholder="Enter input context here", show_label=True)
+    #         with gr.Column():
+    #             saved_llm_outputs = gr.State(value={})
+    #             with gr.Group():
+    #                 selected_base_llm_name_dropdown = gr.Dropdown(label="Base LLM",
+    #                     choices=[f"LLM-{i+1}" for i in range(MIN_BASE_LLM_NUM)], value="LLM-1", show_label=True)
+    #                 selected_base_llm_output = gr.Textbox(lines=4, label="LLM-1 (Click Save to save current content)",
+    #                     placeholder="Enter LLM-1 output here", show_label=True)
+    #             with gr.Row():
+    #                 base_llm_outputs_save_button = gr.Button('Save', variant='primary')
                     
-                    base_llm_outputs_clear_single_button = gr.Button('Clear Single', variant='primary')
+    #                 base_llm_outputs_clear_single_button = gr.Button('Clear Single', variant='primary')
                     
-                    base_llm_outputs_clear_all_button = gr.Button('Clear All', variant='primary')
-                base_llms_num = gr.Slider(
-                        label='Number of base llms',
-                        minimum=MIN_BASE_LLM_NUM,
-                        maximum=MAX_BASE_LLM_NUM,
-                        step=1,
-                        value=MIN_BASE_LLM_NUM,
-                    )
+    #                 base_llm_outputs_clear_all_button = gr.Button('Clear All', variant='primary')
+    #             base_llms_num = gr.Slider(
+    #                     label='Number of base llms',
+    #                     minimum=MIN_BASE_LLM_NUM,
+    #                     maximum=MAX_BASE_LLM_NUM,
+    #                     step=1,
+    #                     value=MIN_BASE_LLM_NUM,
+    #                 )
         
-        blender_state = gr.State(value={})
-        saved_rank_outputs = gr.State(value=[])
-        saved_fuse_outputs = gr.State(value=[])
-        gr.Markdown("## Blender Outputs")
-        with gr.Group():
-            rank_outputs = gr.Textbox(lines=1, label="Ranking outputs", placeholder="Ranking outputs", show_label=True)
-            fuser_outputs = gr.Textbox(lines=4, label="Fusing outputs", placeholder="Fusing outputs", show_label=True)
-        with gr.Row():
-            rank_button = gr.Button('Rank LLM Outputs', variant='primary')
-            fuse_button = gr.Button('Fuse Top-K ranked outputs', variant='primary')
-            clear_button = gr.Button('Clear Blender Outputs', variant='primary')
-        blender_config = gr.State(value={
-            "source_max_length": DEFAULT_SOURCE_MAX_LENGTH,
-            "candidate_max_length": DEFAULT_CANDIDATE_MAX_LENGTH,
-            "top_k_for_fuser": 3,
-            "max_new_tokens": DEFAULT_FUSER_MAX_NEW_TOKENS,
-            "temperature": 0.7,
-            "top_p": 1.0,
-        })
+    #     blender_state = gr.State(value={})
+    #     saved_rank_outputs = gr.State(value=[])
+    #     saved_fuse_outputs = gr.State(value=[])
+    #     gr.Markdown("## Blender Outputs")
+    #     with gr.Group():
+    #         rank_outputs = gr.Textbox(lines=1, label="Ranking outputs", placeholder="Ranking outputs", show_label=True)
+    #         fuser_outputs = gr.Textbox(lines=4, label="Fusing outputs", placeholder="Fusing outputs", show_label=True)
+    #     with gr.Row():
+    #         rank_button = gr.Button('Rank LLM Outputs', variant='primary')
+    #         fuse_button = gr.Button('Fuse Top-K ranked outputs', variant='primary')
+    #         clear_button = gr.Button('Clear Blender Outputs', variant='primary')
+    #     blender_config = gr.State(value={
+    #         "source_max_length": DEFAULT_SOURCE_MAX_LENGTH,
+    #         "candidate_max_length": DEFAULT_CANDIDATE_MAX_LENGTH,
+    #         "top_k_for_fuser": 3,
+    #         "max_new_tokens": DEFAULT_FUSER_MAX_NEW_TOKENS,
+    #         "temperature": 0.7,
+    #         "top_p": 1.0,
+    #     })
             
-        with gr.Accordion(label='Advanced options', open=False):
-            source_max_length = gr.Slider(
-                label='Max length of Instruction + Input',
-                minimum=1,
-                maximum=SOURCE_MAX_LENGTH,
-                step=1,
-                value=DEFAULT_SOURCE_MAX_LENGTH,
-            )
-            candidate_max_length = gr.Slider(
-                label='Max length of LLM-Output Candidate',
-                minimum=1,
-                maximum=CANDIDATE_MAX_LENGTH,
-                step=1,
-                value=DEFAULT_CANDIDATE_MAX_LENGTH,
-            )
-            top_k_for_fuser = gr.Slider(
-                label='Top-k ranked candidates to fuse',
-                minimum=1,
-                maximum=3,
-                step=1,
-                value=3,
-            )
-            max_new_tokens = gr.Slider(
-                label='Max new tokens fuser can generate',
-                minimum=1,
-                maximum=FUSER_MAX_NEW_TOKENS,
-                step=1,
-                value=DEFAULT_FUSER_MAX_NEW_TOKENS,
-            )
-            temperature = gr.Slider(
-                label='Temperature of fuser generation',
-                minimum=0.1,
-                maximum=2.0,
-                step=0.1,
-                value=0.7,
-            )
-            top_p = gr.Slider(
-                label='Top-p of fuser generation',
-                minimum=0.05,
-                maximum=1.0,
-                step=0.05,
-                value=1.0,
-            )
+    #     with gr.Accordion(label='Advanced options', open=False):
+    #         source_max_length = gr.Slider(
+    #             label='Max length of Instruction + Input',
+    #             minimum=1,
+    #             maximum=SOURCE_MAX_LENGTH,
+    #             step=1,
+    #             value=DEFAULT_SOURCE_MAX_LENGTH,
+    #         )
+    #         candidate_max_length = gr.Slider(
+    #             label='Max length of LLM-Output Candidate',
+    #             minimum=1,
+    #             maximum=CANDIDATE_MAX_LENGTH,
+    #             step=1,
+    #             value=DEFAULT_CANDIDATE_MAX_LENGTH,
+    #         )
+    #         top_k_for_fuser = gr.Slider(
+    #             label='Top-k ranked candidates to fuse',
+    #             minimum=1,
+    #             maximum=3,
+    #             step=1,
+    #             value=3,
+    #         )
+    #         max_new_tokens = gr.Slider(
+    #             label='Max new tokens fuser can generate',
+    #             minimum=1,
+    #             maximum=FUSER_MAX_NEW_TOKENS,
+    #             step=1,
+    #             value=DEFAULT_FUSER_MAX_NEW_TOKENS,
+    #         )
+    #         temperature = gr.Slider(
+    #             label='Temperature of fuser generation',
+    #             minimum=0.1,
+    #             maximum=2.0,
+    #             step=0.1,
+    #             value=0.7,
+    #         )
+    #         top_p = gr.Slider(
+    #             label='Top-p of fuser generation',
+    #             minimum=0.05,
+    #             maximum=1.0,
+    #             step=0.05,
+    #             value=1.0,
+    #         )
         
-        examples_dummy_textbox = gr.Textbox(lines=1, label="", placeholder="", show_label=False, visible=False)     
-        batch_examples = gr.Examples(
-            examples=MIX_INSTRUCT_EXAMPLES,
-            fn=get_preprocess_examples,
-            cache_examples=True,
-            examples_per_page=5,
-            inputs=[inst_textbox, input_textbox],
-            outputs=[inst_textbox, input_textbox, base_llms_num, examples_dummy_textbox],
-        )
+    #     examples_dummy_textbox = gr.Textbox(lines=1, label="", placeholder="", show_label=False, visible=False)     
+    #     batch_examples = gr.Examples(
+    #         examples=MIX_INSTRUCT_EXAMPLES,
+    #         fn=get_preprocess_examples,
+    #         cache_examples=True,
+    #         examples_per_page=5,
+    #         inputs=[inst_textbox, input_textbox],
+    #         outputs=[inst_textbox, input_textbox, base_llms_num, examples_dummy_textbox],
+    #     )
             
-        base_llms_num.input(
-            fn=update_base_llms_num,
-            inputs=[base_llms_num, saved_llm_outputs],
-            outputs=[selected_base_llm_name_dropdown, saved_llm_outputs],
-        )
+    #     base_llms_num.input(
+    #         fn=update_base_llms_num,
+    #         inputs=[base_llms_num, saved_llm_outputs],
+    #         outputs=[selected_base_llm_name_dropdown, saved_llm_outputs],
+    #     )
         
-        examples_dummy_textbox.change(
-            fn=update_base_llm_dropdown_along_examples,
-            inputs=[inst_textbox, input_textbox],
-            outputs=[saved_llm_outputs, rank_outputs, fuser_outputs, selected_base_llm_name_dropdown],
-        ).then(
-            fn=display_llm_output,
-            inputs=[saved_llm_outputs, selected_base_llm_name_dropdown],
-            outputs=selected_base_llm_output,
-        )
+    #     examples_dummy_textbox.change(
+    #         fn=update_base_llm_dropdown_along_examples,
+    #         inputs=[inst_textbox, input_textbox],
+    #         outputs=[saved_llm_outputs, rank_outputs, fuser_outputs, selected_base_llm_name_dropdown],
+    #     ).then(
+    #         fn=display_llm_output,
+    #         inputs=[saved_llm_outputs, selected_base_llm_name_dropdown],
+    #         outputs=selected_base_llm_output,
+    #     )
         
-        selected_base_llm_name_dropdown.change(
-            fn=display_llm_output,
-            inputs=[saved_llm_outputs, selected_base_llm_name_dropdown],
-            outputs=selected_base_llm_output,
-        )
+    #     selected_base_llm_name_dropdown.change(
+    #         fn=display_llm_output,
+    #         inputs=[saved_llm_outputs, selected_base_llm_name_dropdown],
+    #         outputs=selected_base_llm_output,
+    #     )
         
-        base_llm_outputs_save_button.click(
-            fn=save_llm_output,
-            inputs=[selected_base_llm_name_dropdown, selected_base_llm_output, saved_llm_outputs],
-            outputs=saved_llm_outputs,
-        )
-        base_llm_outputs_clear_all_button.click(
-            fn=lambda: [{}, ""],
-            inputs=[],
-            outputs=[saved_llm_outputs, selected_base_llm_output],
-        )
-        base_llm_outputs_clear_single_button.click(
-            fn=lambda: "",
-            inputs=[],
-            outputs=selected_base_llm_output,
-        )
+    #     base_llm_outputs_save_button.click(
+    #         fn=save_llm_output,
+    #         inputs=[selected_base_llm_name_dropdown, selected_base_llm_output, saved_llm_outputs],
+    #         outputs=saved_llm_outputs,
+    #     )
+    #     base_llm_outputs_clear_all_button.click(
+    #         fn=lambda: [{}, ""],
+    #         inputs=[],
+    #         outputs=[saved_llm_outputs, selected_base_llm_output],
+    #     )
+    #     base_llm_outputs_clear_single_button.click(
+    #         fn=lambda: "",
+    #         inputs=[],
+    #         outputs=selected_base_llm_output,
+    #     )
             
 
-        rank_button.click(
-            fn=check_save_ranker_inputs,
-            inputs=[inst_textbox, input_textbox, saved_llm_outputs, blender_config],
-            outputs=blender_state,
-        ).success(
-            fn=llms_rank,
-            inputs=[inst_textbox, input_textbox, saved_llm_outputs, blender_config],
-            outputs=[saved_rank_outputs, rank_outputs],
-        )
+    #     rank_button.click(
+    #         fn=check_save_ranker_inputs,
+    #         inputs=[inst_textbox, input_textbox, saved_llm_outputs, blender_config],
+    #         outputs=blender_state,
+    #     ).success(
+    #         fn=llms_rank,
+    #         inputs=[inst_textbox, input_textbox, saved_llm_outputs, blender_config],
+    #         outputs=[saved_rank_outputs, rank_outputs],
+    #     )
         
-        fuse_button.click(
-            fn=check_fuser_inputs,
-            inputs=[blender_state, blender_config, saved_rank_outputs],
-            outputs=[],
-        ).success(
-            fn=llms_fuse,
-            inputs=[blender_state, blender_config, saved_rank_outputs],
-            outputs=[saved_fuse_outputs, fuser_outputs],
-        )
+    #     fuse_button.click(
+    #         fn=check_fuser_inputs,
+    #         inputs=[blender_state, blender_config, saved_rank_outputs],
+    #         outputs=[],
+    #     ).success(
+    #         fn=llms_fuse,
+    #         inputs=[blender_state, blender_config, saved_rank_outputs],
+    #         outputs=[saved_fuse_outputs, fuser_outputs],
+    #     )
         
-        clear_button.click(
-            fn=lambda: ["", "", {}, []],
-            inputs=[],
-            outputs=[rank_outputs, fuser_outputs, blender_state, saved_rank_outputs],
-        )
+    #     clear_button.click(
+    #         fn=lambda: ["", "", {}, []],
+    #         inputs=[],
+    #         outputs=[rank_outputs, fuser_outputs, blender_state, saved_rank_outputs],
+    #     )
         
-        # update blender config
-        source_max_length.change(
-            fn=lambda x, y: y.update({"source_max_length": x}) or y,
-            inputs=[source_max_length, blender_config],
-            outputs=blender_config,
-        )
-        candidate_max_length.change(
-            fn=lambda x, y: y.update({"candidate_max_length": x}) or y,
-            inputs=[candidate_max_length, blender_config],
-            outputs=blender_config,
-        )
-        top_k_for_fuser.change(
-            fn=lambda x, y: y.update({"top_k_for_fuser": x}) or y,
-            inputs=[top_k_for_fuser, blender_config],
-            outputs=blender_config,
-        )
-        max_new_tokens.change(
-            fn=lambda x, y: y.update({"max_new_tokens": x}) or y,
-            inputs=[max_new_tokens, blender_config],
-            outputs=blender_config,
-        )
-        temperature.change(
-            fn=lambda x, y: y.update({"temperature": x}) or y,
-            inputs=[temperature, blender_config],
-            outputs=blender_config,
-        )
-        top_p.change(
-            fn=lambda x, y: y.update({"top_p": x}) or y,
-            inputs=[top_p, blender_config],
-            outputs=blender_config,
-        )
+    #     # update blender config
+    #     source_max_length.change(
+    #         fn=lambda x, y: y.update({"source_max_length": x}) or y,
+    #         inputs=[source_max_length, blender_config],
+    #         outputs=blender_config,
+    #     )
+    #     candidate_max_length.change(
+    #         fn=lambda x, y: y.update({"candidate_max_length": x}) or y,
+    #         inputs=[candidate_max_length, blender_config],
+    #         outputs=blender_config,
+    #     )
+    #     top_k_for_fuser.change(
+    #         fn=lambda x, y: y.update({"top_k_for_fuser": x}) or y,
+    #         inputs=[top_k_for_fuser, blender_config],
+    #         outputs=blender_config,
+    #     )
+    #     max_new_tokens.change(
+    #         fn=lambda x, y: y.update({"max_new_tokens": x}) or y,
+    #         inputs=[max_new_tokens, blender_config],
+    #         outputs=blender_config,
+    #     )
+    #     temperature.change(
+    #         fn=lambda x, y: y.update({"temperature": x}) or y,
+    #         inputs=[temperature, blender_config],
+    #         outputs=blender_config,
+    #     )
+    #     top_p.change(
+    #         fn=lambda x, y: y.update({"top_p": x}) or y,
+    #         inputs=[top_p, blender_config],
+    #         outputs=blender_config,
+    #     )