Spaces:
Build error
Build error
Commit
·
05a5906
1
Parent(s):
c7de84b
Add function to retrieve chatbot rankings
Browse filesA new function called 'get_rankings' has been added to the 'guardrails_buttons.py' file. This function fetches rankings from an API and generates a sorted markdown table. Also, necessary changes have been made to other files to integrate and display this ranking data in the user interface.
- app.py +7 -9
- guardrails_buttons.py +27 -0
- guardrails_models.py +2 -2
- llamaguard_moderator.py +0 -2
app.py
CHANGED
|
@@ -13,6 +13,7 @@ from guardrails_buttons import (
|
|
| 13 |
deactivate_invisible_vote_buttons,
|
| 14 |
deactivate_textbox,
|
| 15 |
deactivate_visible_vote_buttons,
|
|
|
|
| 16 |
leftvote,
|
| 17 |
rightvote,
|
| 18 |
share_js,
|
|
@@ -213,7 +214,7 @@ with gr.Blocks(
|
|
| 213 |
background-size: cover; /* Adjust as needed */
|
| 214 |
background-position: center;
|
| 215 |
}
|
| 216 |
-
#model_description_markdown table {
|
| 217 |
width: 100%;
|
| 218 |
}
|
| 219 |
.w-100 {
|
|
@@ -280,6 +281,7 @@ with gr.Blocks(
|
|
| 280 |
states = [gr.State() for _ in range(num_sides)]
|
| 281 |
chatbots = [None] * num_sides
|
| 282 |
models = gr.State(get_random_models)
|
|
|
|
| 283 |
system_prompt = gr.State(get_random_system_prompt)
|
| 284 |
show_models = [None] * num_sides
|
| 285 |
conversation_id = gr.State()
|
|
@@ -447,14 +449,10 @@ with gr.Blocks(
|
|
| 447 |
"""
|
| 448 |
)
|
| 449 |
|
| 450 |
-
with gr.Tab(label="🏆 Leaderboard"):
|
| 451 |
-
gr.Markdown(
|
| 452 |
-
|
| 453 |
-
|
| 454 |
-
|
| 455 |
-
We will launch the guardrails leaderboard once enough votes are collected. Ranking will be calculated based on ELO ratings. Keep playing so that we can collect enough data.
|
| 456 |
-
"""
|
| 457 |
-
)
|
| 458 |
|
| 459 |
gr.Markdown(
|
| 460 |
"""
|
|
|
|
| 13 |
deactivate_invisible_vote_buttons,
|
| 14 |
deactivate_textbox,
|
| 15 |
deactivate_visible_vote_buttons,
|
| 16 |
+
get_rankings,
|
| 17 |
leftvote,
|
| 18 |
rightvote,
|
| 19 |
share_js,
|
|
|
|
| 214 |
background-size: cover; /* Adjust as needed */
|
| 215 |
background-position: center;
|
| 216 |
}
|
| 217 |
+
#model_description_markdown table, #leaderboard table {
|
| 218 |
width: 100%;
|
| 219 |
}
|
| 220 |
.w-100 {
|
|
|
|
| 281 |
states = [gr.State() for _ in range(num_sides)]
|
| 282 |
chatbots = [None] * num_sides
|
| 283 |
models = gr.State(get_random_models)
|
| 284 |
+
rankings = gr.State("")
|
| 285 |
system_prompt = gr.State(get_random_system_prompt)
|
| 286 |
show_models = [None] * num_sides
|
| 287 |
conversation_id = gr.State()
|
|
|
|
| 449 |
"""
|
| 450 |
)
|
| 451 |
|
| 452 |
+
with gr.Tab(label="🏆 Leaderboard", elem_id="leaderboard") as leaderboard_tab:
|
| 453 |
+
gr.Markdown("## 🏆 Guardrails Leaderboard")
|
| 454 |
+
rankings = gr.Markdown("")
|
| 455 |
+
leaderboard_tab.select(get_rankings, None, [rankings])
|
|
|
|
|
|
|
|
|
|
|
|
|
| 456 |
|
| 457 |
gr.Markdown(
|
| 458 |
"""
|
guardrails_buttons.py
CHANGED
|
@@ -143,6 +143,33 @@ def bothbadvote(conversation_id, history1, history2):
|
|
| 143 |
pass
|
| 144 |
|
| 145 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 146 |
def show_models_fn(models):
|
| 147 |
model_1 = gr.Markdown(" 🅰️ " + models[0]["name"])
|
| 148 |
model_2 = gr.Markdown(" 🅱️ " + models[1]["name"])
|
|
|
|
| 143 |
pass
|
| 144 |
|
| 145 |
|
| 146 |
+
def get_rankings():
|
| 147 |
+
try:
|
| 148 |
+
response = requests.get(f"{LIGHTHOUZ_API_URL}/rankings")
|
| 149 |
+
ratings = response.json()["ratings"]
|
| 150 |
+
sorted_ratings = dict(
|
| 151 |
+
sorted(ratings.items(), key=lambda item: item[1], reverse=True)
|
| 152 |
+
)
|
| 153 |
+
|
| 154 |
+
markdown_table = "| Rank | Chatbots | Arena Elo |\n|:-----:|-----|-------|\n"
|
| 155 |
+
current_rank = 1
|
| 156 |
+
last_value = None
|
| 157 |
+
rank_increment = 0
|
| 158 |
+
for key, value in sorted_ratings.items():
|
| 159 |
+
rounded_value = round(value)
|
| 160 |
+
if last_value is None or rounded_value != last_value:
|
| 161 |
+
current_rank += rank_increment
|
| 162 |
+
rank_increment = 0
|
| 163 |
+
markdown_table += f"| {current_rank} | {key} | {rounded_value} |\n"
|
| 164 |
+
last_value = rounded_value
|
| 165 |
+
rank_increment += 1
|
| 166 |
+
|
| 167 |
+
return markdown_table
|
| 168 |
+
except Exception as e:
|
| 169 |
+
print(e)
|
| 170 |
+
return ""
|
| 171 |
+
|
| 172 |
+
|
| 173 |
def show_models_fn(models):
|
| 174 |
model_1 = gr.Markdown(" 🅰️ " + models[0]["name"])
|
| 175 |
model_2 = gr.Markdown(" 🅱️ " + models[1]["name"])
|
guardrails_models.py
CHANGED
|
@@ -5,9 +5,9 @@ from typing import List, Optional
|
|
| 5 |
import openai
|
| 6 |
from google.generativeai.types import (
|
| 7 |
BlockedPromptException,
|
| 8 |
-
StopCandidateException,
|
| 9 |
-
HarmCategory,
|
| 10 |
HarmBlockThreshold,
|
|
|
|
|
|
|
| 11 |
)
|
| 12 |
from langchain_community.chat_models import ChatAnyscale
|
| 13 |
from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
|
|
|
|
| 5 |
import openai
|
| 6 |
from google.generativeai.types import (
|
| 7 |
BlockedPromptException,
|
|
|
|
|
|
|
| 8 |
HarmBlockThreshold,
|
| 9 |
+
HarmCategory,
|
| 10 |
+
StopCandidateException,
|
| 11 |
)
|
| 12 |
from langchain_community.chat_models import ChatAnyscale
|
| 13 |
from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
|
llamaguard_moderator.py
CHANGED
|
@@ -127,8 +127,6 @@ Provide your safety assessment for $agent_type in the above conversation:
|
|
| 127 |
- If unsafe, a second line must include a comma-separated list of violated categories."""
|
| 128 |
)
|
| 129 |
|
| 130 |
-
print(PROMPT_TEMPLATE)
|
| 131 |
-
|
| 132 |
|
| 133 |
def format_prompt(role, message):
|
| 134 |
prompt = PROMPT_TEMPLATE.substitute(prompt=message, agent_type=role)
|
|
|
|
| 127 |
- If unsafe, a second line must include a comma-separated list of violated categories."""
|
| 128 |
)
|
| 129 |
|
|
|
|
|
|
|
| 130 |
|
| 131 |
def format_prompt(role, message):
|
| 132 |
prompt = PROMPT_TEMPLATE.substitute(prompt=message, agent_type=role)
|