File size: 12,106 Bytes
31fda98
061ae36
31fda98
06d4d61
37aa22e
31fda98
06826f1
 
81602d3
31fda98
06826f1
061ae36
06826f1
061ae36
06826f1
 
 
 
 
 
06d4d61
 
061ae36
 
06d4d61
 
 
 
 
 
 
 
 
06826f1
 
 
 
06d4d61
31fda98
06d4d61
 
31fda98
 
06d4d61
 
 
 
 
 
31fda98
 
 
 
 
06d4d61
 
 
31fda98
 
 
 
 
 
 
 
 
06d4d61
31fda98
06d4d61
31fda98
06d4d61
 
31fda98
 
06d4d61
81602d3
 
b9ee4b2
 
81602d3
 
 
 
 
 
 
b9ee4b2
81602d3
 
 
 
 
 
 
31fda98
 
06826f1
81602d3
b9ee4b2
 
 
 
 
 
81602d3
 
b9ee4b2
 
 
81602d3
b9ee4b2
81602d3
31fda98
 
b9ee4b2
 
81602d3
 
 
 
06d4d61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
31fda98
81602d3
 
b9ee4b2
 
06d4d61
 
81602d3
06d4d61
81602d3
06d4d61
81602d3
b9ee4b2
81602d3
b9ee4b2
 
 
81602d3
 
 
 
 
06d4d61
81602d3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
06d4d61
31fda98
06d4d61
b9ee4b2
06d4d61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
31fda98
 
06d4d61
31fda98
 
06d4d61
 
 
 
 
 
 
 
 
 
 
 
 
81602d3
 
 
06d4d61
 
31fda98
06d4d61
 
 
 
 
 
 
81602d3
 
 
31fda98
 
81602d3
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
import gradio as gr
import spaces
import pandas as pd
from typing import List, Dict, Tuple
from flow_judge import Hf, FlowJudge, EvalInput
from flow_judge.metrics import CustomMetric, RubricItem
from huggingface_hub import snapshot_download
from flow_judge.models.huggingface import Hf
from examples import get_examples

MODEL_NAME = "flowaicom/Flow-Judge-v0.1"

def download_model():
    try:
        print(f"Downloading model {MODEL_NAME}...")
        snapshot_download(repo_id=MODEL_NAME)
        print(f"Model {MODEL_NAME} downloaded to default Hugging Face cache")
        return True
    except Exception as e:
        raise RuntimeError(f"Failed to download model {MODEL_NAME}: {e}")
    
    
    
@spaces.GPU
def evaluate(
    inputs_task: List[Dict[str, str]],
    output_name: str,
    output_value: str,
    evaluation_criteria: str,
    rubric_items: List[Dict[str, str]]
) -> Tuple[str, int]:
    
    # [{'name': 'a', 'value': 'a'}]
    try:
        model = Hf(flash_attn=False)
    except Exception as e:
        raise RuntimeError(f"Failed to initialize Hf Model: {e}")
    
    eval_input = EvalInput(
        inputs=[{input['name']: input['value']} for input in inputs_task],
        output={output_name: output_value}
    )
    
    score_rubric_items = [
        RubricItem(
            score=int(rubric_item['name']), 
            description=rubric_item['value']
        )
        for rubric_item in rubric_items
    ]
    
    custom_metric = CustomMetric(
        name="custom-metric",
        criteria=evaluation_criteria,
        rubric=score_rubric_items,
        required_inputs=[input['name'] for input in inputs_task],
        required_output=output_name
    )
    
    judge = FlowJudge(model=model, metric=custom_metric)
    
    try:
        result = judge.evaluate(eval_input)
    except Exception as e:
        raise RuntimeError(f"Failed to evaluate: {e}")
    
    return result.feedback, result.score

def reset_all():
    return (
        [], "", "", [], "", "",  # Existing resets for inputs and rubrics
        "", "", "", "", ""      # New resets for additional fields
    )

# Define presets
EXAMPLES = get_examples()

IMAGE_PATH = "./img/flow_judge_banner.png"

HEADER = """<h1 align="center" style="font-family: 'Courier New', Courier, monospace;">Flow Judge Demo</h1>

<p align="center" style="font-family: 'Courier New', Courier, monospace;">
  <strong>
    <a href="https://www.flow-ai.com/judge">Technical Report</a> |
    <a href="https://huggingface.co/collections/flowaicom/flow-judge-v01-66e6af5fc3b3a128bde07dec">Model Weights</a> |
    <a href="https://github.com/flowaicom/lm-evaluation-harness/tree/Flow-Judge-v0.1_evals/lm_eval/tasks/flow_judge_evals">Evaluation Code</a> |
    <a href="https://github.com/flowaicom/flow-judge/tree/main/examples">Tutorials</a>
  </strong>
</p>

<p align="center" style="font-family: 'Courier New', Courier, monospace;">
  <code>flow-judge</code> is a lightweight library for evaluating LLM applications with <code>Flow-Judge-v0.1</code>.
</p>"""


with gr.Blocks() as demo:
    model_downloaded = download_model()
    
    with gr.Row(equal_height=False):
        with gr.Column(scale=2):
            gr.Image(IMAGE_PATH, show_label=False, interactive=False, show_share_button=False, show_fullscreen_button=False, show_download_button=False)
        with gr.Column(scale=3):
            gr.HTML(HEADER)
    gr.Markdown("# ⚡ **Quickstart Examples**")
    with gr.Row():
        with gr.Column(scale=1):
            preset_buttons = [gr.Button(example["description"]) for example in EXAMPLES[:len(EXAMPLES)//3]]
        with gr.Column(scale=1):
            preset_buttons += [gr.Button(example["description"]) for example in EXAMPLES[len(EXAMPLES)//3:2*len(EXAMPLES)//3]]
        with gr.Column(scale=1):
            preset_buttons += [gr.Button(example["description"]) for example in EXAMPLES[2*len(EXAMPLES)//3:]]
    
    with gr.Row(equal_height=False):
        with gr.Column(scale=1):
            gr.Markdown("## **Evaluation task inputs**")
            gr.Markdown("*<span style='color: gray;'>Define the input names and values for the evaluation task. Inputs are optional if evaluation depends on the output only.</span>*")
            with gr.Group():
                inputs_task = gr.State([])
                new_input_name = gr.Textbox(label="Name")
                new_input_value = gr.Textbox(label="Value")
        
            def add_input(inputs_task, new_input_name, new_input_value):
                return inputs_task + [{"name": new_input_name, "value": new_input_value}], "", ""
            
            @gr.render(inputs=inputs_task) # You have to pass the state here
            def render_inputs(inputs_list): # Use different name than the state variable
                
                for input in inputs_list:
                    with gr.Group():
                        with gr.Row(equal_height=True):
                            with gr.Column(min_width=60, scale=2):
                                gr.Textbox(input['name'], label="Name", show_label=True, interactive=False, autoscroll=False, max_lines=1)
                            with gr.Column(scale=8):
                                gr.Textbox(input['value'], label="Value", show_label=True, interactive=False, autoscroll=False, max_lines=3)
                            with gr.Column(min_width=15, scale=1):
                                delete_btn = gr.Button("X", size="lg", variant="secondary")
                                def delete(input=input):
                                    inputs_list.remove(input)
                                    return inputs_list
                            delete_btn.click(delete, None, [inputs_task]) # This is the state variable
                        
            gr.Button("Add Input").click(
                add_input,
                [inputs_task, new_input_name, new_input_value],
                [inputs_task, new_input_name, new_input_value]
            )
        
        with gr.Column(scale=1):
            gr.Markdown("## **Evaluation task output**")
            gr.Markdown("*<span style='color: gray;'>Define the output name and value for the evaluation task. Output is always required.</span>*")
            with gr.Group():
                with gr.Row(equal_height=True):
                    with gr.Column(min_width=60, scale=2):
                        output_name = gr.Textbox(label="Name", show_label=True, interactive=True, autoscroll=False, max_lines=1)
                    with gr.Column(scale=9):
                        output_value = gr.Textbox(label="Value", show_label=True, interactive=True, autoscroll=False, max_lines=3)
    
   
    with gr.Column(scale=1):
        gr.Markdown("## **Evaluation criteria and rubric**")
        gr.Markdown("*<span style='color: gray;'>Define the evaluation criteria and rubric for the evaluation task. Supported scoring scales: Binary (0 and 1), 3-Likert and 5-Likert.</span>*\n\n*<span style='color: gray;'>❗You can experiment with other scoring scales. However, performance may vary.</span>*")

        with gr.Row():
            with gr.Column(scale=1):
                rubric_items = gr.State([])
                new_rubric_name = gr.Textbox(label="Score", show_label=True, interactive=True, autoscroll=False, max_lines=1)
                new_rubric_value = gr.Textbox(label="Description", show_label=True, interactive=True, autoscroll=False, max_lines=3)
                
                def add_rubric_item(rubric_items, new_rubric_name, new_rubric_value):
                    return rubric_items + [{"name": new_rubric_name, "value": new_rubric_value}], "", ""
                
                @gr.render(inputs=rubric_items) # You have to pass the state here
                def render_rubrics(rubric_items_list): # Use different name than the state variable
                    
                    for rubric_item in rubric_items_list:
                        with gr.Group():
                            with gr.Row(equal_height=True):
                                with gr.Column(min_width=30, scale=1):
                                    gr.Textbox(rubric_item['name'], label="Score", show_label=True, interactive=False)
                                with gr.Column(scale=8):
                                    gr.Textbox(rubric_item['value'], label="Description", show_label=True, interactive=False)
                                with gr.Column(min_width=15, scale=1):
                                    delete_btn = gr.Button("X", size="lg", variant="secondary")
                                    def delete(rubric_item=rubric_item):
                                        rubric_items_list.remove(rubric_item)
                                        return rubric_items_list
                                    delete_btn.click(delete, None, [rubric_items]) # This is the state variable
                                
                gr.Button("Add Rubric Item").click(
                    add_rubric_item,
                    [rubric_items, new_rubric_name, new_rubric_value],
                    [rubric_items, new_rubric_name, new_rubric_value]
                )
            with gr.Column(scale=1):
                evaluation_criteria = gr.Textbox(label="Evaluation criteria")
                            
    with gr.Row():
        with gr.Column(scale=1, variant="panel"):
            gr.Markdown("# **Evaluation**")
            with gr.Group():
                with gr.Row(equal_height=True):
                    with gr.Column(min_width=15, scale=1):
                        score = gr.Textbox(label="Score", interactive=False, autoscroll=False, max_lines=1)
                    with gr.Column(scale=5):
                        feedback = gr.Textbox(label="Feedback", interactive=False, autoscroll=False, max_lines=6)
                    with gr.Column(min_width=15, scale=1):
                        evaluate_btn = gr.Button("Evaluate", variant="primary")

                        reset_all_btn = gr.Button("Clear All", variant="stop")  # Add Reset All button
                        reset_all_btn.click(
                            reset_all,
                            inputs=[], 
                            outputs=[
                                inputs_task, 
                                new_input_name, 
                                new_input_value, 
                                rubric_items, 
                                new_rubric_name, 
                                new_rubric_value,
                                evaluation_criteria,    # Reset evaluation criteria
                                output_name,            # Reset output name
                                output_value,           # Reset output value
                                feedback,               # Reset feedback
                                score                   # Reset score
                            ]
                        )
    
    evaluate_btn.click(
        evaluate,
        inputs=[inputs_task, output_name, output_value, evaluation_criteria, rubric_items],
        outputs=[feedback, score]
    )
    
    for i, button in enumerate(preset_buttons):
        def populate_preset(ex_i=i):
            return populate_fields(ex_i)
    
        button.click(
            populate_preset,  # Use the closure to pass the current index
            inputs=[],  # No direct inputs needed
            outputs=[
                inputs_task, 
                output_name, 
                output_value, 
                evaluation_criteria, 
                rubric_items,
                feedback,  # Add feedback to be reset
                score  # Add score to be reset
            ]
        )

def populate_fields(example_index: int):
    example = EXAMPLES[example_index]
    return (
        example["inputs_task"],
        example["output"]["name"],
        example["output"]["value"],
        example["evaluation_criteria"],
        example["rubric"],
        "",  # Reset feedback
        ""  # Reset score
    )

demo.launch(debug=True)