File size: 10,253 Bytes
0467ec9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
import streamlit as st
from PIL import Image
import codecs
import streamlit.components.v1 as components
from utils import inject_custom_css
import streamlit as st
from streamlit_plotly_events import plotly_events
import pickle
import matplotlib.pyplot as plt
import plotly.graph_objects as go
import typing as tp
import colorsys

plt.style.use('default')
plt.rcParams['text.usetex'] = True
plt.rcParams['font.family'] = 'serif'


def interpolate_color(color1, color2, factor):
    """Interpolates between two RGB colors. Factor is between 0 and 1."""
    color1 = colorsys.rgb_to_hls(
        int(color1[1:3], 16) / 255.0,
        int(color1[3:5], 16) / 255.0,
        int(color1[5:], 16) / 255.0
    )
    color2 = colorsys.rgb_to_hls(
        int(color2[1:3], 16) / 255.0,
        int(color2[3:5], 16) / 255.0,
        int(color2[5:], 16) / 255.0
    )
    new_color = [color1[i] * (1 - factor) + color2[i] * factor for i in range(3)]
    new_color = colorsys.hls_to_rgb(*new_color)
    return '#{:02x}{:02x}{:02x}'.format(
        int(new_color[0] * 255), int(new_color[1] * 255), int(new_color[2] * 255)
    )


color1 = "#fa7659"
color2 = "#6dafd7"

shapes = [
    dict(
        type="rect",
        xref="paper",
        yref="paper",
        x0=0,
        y0=0,
        x1=1,
        y1=1,
        line=dict(
            color="Black",
            width=2,
        ),
    )
]

shapes = [
    dict(
        type="rect",
        xref="paper",
        yref="paper",
        x0=0,
        y0=0,
        x1=1,
        y1=1,
        line=dict(
            color="Black",
            width=2,
        ),
    )
]


def plot_pareto(dict_results: tp.Dict):

    reward1_key = "R1"
    reward2_key = "R2"

    # Series for "wa"
    dict_results["wa_d"] = [x for i, x in enumerate(dict_results["wa_d"]) if i % 2 == 0]
    lambda_values_wa = [
        round(i / (len(dict_results["wa_d"]) - 1), 2) for i in range(len(dict_results["wa_d"]))
    ][::-1]
    reward1_values_wa = [item[reward1_key] for item in dict_results["wa_d"]]
    reward2_values_wa = [item[reward2_key] for item in dict_results["wa_d"]]

    # Series for "morl"
    # Series for "init"
    reward1_values_morl = [dict_results["morl"][reward1_key]]
    reward2_values_morl = [dict_results["morl"][reward2_key]]

    # Series for "init"
    reward1_values_init = [dict_results["init"][reward1_key]]
    reward2_values_init = [dict_results["init"][reward2_key]]

    layout = go.Layout(autosize=False, width=1000, height=1000)
    fig = go.Figure(layout=layout)

    for i in range(len(reward1_values_wa) - 1):
        fig.add_trace(
            go.Scatter(
                x=reward1_values_wa[i:i + 2],
                y=reward2_values_wa[i:i + 2],
                mode='lines',
                hoverinfo='skip',
                line=dict(
                    color=interpolate_color(color1, color2, i / (len(reward1_values_wa) - 1)),
                    width=2
                ),
                showlegend=False
            )
        )

    # Plot for "wa"
    fig.add_trace(
        go.Scatter(
            x=reward1_values_wa,
            y=reward2_values_wa,
            mode='markers',
            name='Rewarded soups: 0≤λ≤1',
            hoverinfo='text',
            hovertext=[f'λ={lmbda}' for lmbda in lambda_values_wa],
            marker=dict(
                color=[
                    interpolate_color(color1, color2, i / len(lambda_values_wa))
                    for i in range(len(lambda_values_wa))
                ],
                size=10
            )
        )
    )

    # Plot for "morl"
    fig.add_trace(
        go.Scatter(
            x=reward1_values_morl,
            y=reward2_values_morl,
            mode='markers',
            name='MORL: μ=0.5',
            hoverinfo='skip',
            marker=dict(color='#A45EE9', size=15, symbol="star"),
        )
    )

    # Plot for "init"
    fig.add_trace(
        go.Scatter(
            x=reward1_values_init,
            y=reward2_values_init,
            mode='markers',
            name='Pre-trained init',
            hoverinfo='skip',
            marker=dict(color='#9f9bc8', size=15, symbol="star"),
        )
    )

    fig.update_layout(
        xaxis=dict(
            #range = [5.21,5.31],
            #nticks=6,
            showticklabels=True,
            ticks='outside',
            tickfont=dict(size=18,),
            title=dict(text="R1", font=dict(size=18), standoff=10),
            showgrid=False,
            zeroline=False,
            hoverformat='.2f'
        ),
        yaxis=dict(
            #range = [0.78,0.825],
            #nticks=7,
            showticklabels=True,
            ticks='outside',
            tickfont=dict(size=18,),
            title=dict(text="R2", font=dict(size=18), standoff=10),
            showgrid=False,
            zeroline=False,
            hoverformat='.2f'
        ),
        font=dict(family="Roboto", size=12, color="Black"),
        hovermode='x unified',
        autosize=False,
        width=500,
        height=500,
        margin=dict(l=100, r=50, b=150, t=20, pad=0),
        paper_bgcolor="White",
        plot_bgcolor="White",
        shapes=shapes,
        legend=dict(
            x=0.5,
            y=0.03,
            traceorder="normal",
            font=dict(family="Roboto", size=12, color="black"),
            bgcolor="White",
            bordercolor="Black",
            borderwidth=1
        )
    )

    return fig


def run():

    st.write(
        f"""
        <link href='http://fonts.googleapis.com/css?family=Roboto' rel='stylesheet' type='text/css'>
        <script src="https://cdn.plot.ly/plotly-latest.min.js"></script>
        <script type="text/javascript" async src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.5/latest.js?config=TeX-MML-AM_CHTML">
        </script>
        <h3 style='text-align: left;';>RLHF of LLaMA for diverse news summarization</h3>""",
        unsafe_allow_html=True
    )

    st.markdown(
        r"""
Given the importance of RLHF to train LLMs, we begin our experiments with text-to-text generation.
Our pre-trained network is LLaMA-7b, instruction fine-tuned on Alpaca.
For RL training with PPO, we employ the trl package and the setup from with low-rank adapters (LoRA) for efficiency.
Here we consider summarization on Reuter news.
To evaluate the summary in the absence of supervision, we utilized two different reward models, available on HuggingFace: [$R_1$](https://huggingface.co./Tristan/gpt2_reward_summarization) follows the Summarize from Human Feedback paper while [$R_2$](https://huggingface.co./CogComp/bart-faithful-summary-detector) leverages contrast candidate generation.

Our results below reveal the following insights. The front defined by rewarded soups between the two weights specialized on $R_1$ (i.e., $\lambda=0.0$) and $R_2$ (i.e., $\lambda=1.0$) is above the straight line connecting those two points; this validates what we call in the paper *the linear mode connectivity hypothesis*. Moreover, the front intersects the point obtained by multi-objective RL (MORL) fine-tuning on $(1-\mu) \times R_1 + \mu \times R_2$ for $\mu=0.5$ (i.e., the average of the two rewards). Interestingly, when we compare both full fronts in the paper, they exhibit qualitatively the same shape. The qualitative visual inspections of the generations show that increasing $\lambda$ leads to shorter but more factual summaries; this is because $R_2$ evaluates faithfulness in priority.""",
        unsafe_allow_html=True
    )
    st.markdown(
        """<h3 style='text-align: center;';>Click on a rewarded soup point on the left and select a subject on the right!</h3>""",
        unsafe_allow_html=True
    )

    files = []

    with open("streamlit_app/data/textgen/data.pkl", "rb") as f:
        data = pickle.load(f)
    with open("streamlit_app/data/textgen/data_prompt.pkl", "rb") as f:
        data_prompt = pickle.load(f)
    with open("streamlit_app/data/textgen/data_title.pkl", "rb") as f:
        data_title = pickle.load(f)

    left, right = st.columns((2, 2))
    with left:
        fig = plot_pareto(data)
        onclick = plotly_events(fig, click_event=True)
    with right:
        option = st.selectbox('', data_title.keys())

        subject = data_title[option]
        st.markdown(
            f"""
        <div class="promptTextbox">
            <div class="promptHeader">
                Text to summarize:
            </div>
            <div class="promptContent">
                {data_prompt[subject]['query']}
            </div>
        </div>
        """,
            unsafe_allow_html=True
        )
        st.markdown("<br>", unsafe_allow_html=True)

        summary1 = data_prompt[subject]['outs'][0]["out"]
        summary3 = data_prompt[subject]['outs'][-1]["out"]
        nb_summaries = len(data_prompt[subject]['outs'])
        if len(onclick) > 0:
            idx = onclick[0]["pointIndex"]
        else:
            idx = 5
        lambda2 = round(1 - idx / (len(data["wa_d"]) - 1), 2)
        summary2 = data_prompt[subject]['outs'][idx]["out"]
        bgcolor = interpolate_color(color2, color1, lambda2)

        st.markdown(
            f"""
        <div class="promptTextbox">
            <div class="promptHeader">
                Generated summaries:
            </div>
            <div class="promptContent">
                <div class="lambda-header" style='background-color: {color2};'>λ=0.0</div><div class="lambdas" style='background-color: {color2};'>{summary3}</div>
                <div class="lambda-header" style='background-color: {bgcolor};'>λ={lambda2} </div><div class="lambdas" style='background-color: {bgcolor};'>{summary2}</div><br>
                <div class="lambda-header" style='background-color: {color1};'>λ=1.0</div><div class="lambdas" style='background-color: {color1};'>{summary1}</div><br>
            </div>
        </div>
        """,
            unsafe_allow_html=True
        )


if __name__ == "__main__":
    img = Image.open("streamlit_app/assets/images/icon.png")
    st.set_page_config(page_title="Rewarded soups", page_icon=img, layout="wide")
    inject_custom_css("streamlit_app/assets/styles.css")
    st.set_option('deprecation.showPyplotGlobalUse', False)
    run()