Spaces:
Sleeping
Sleeping
justinxzhao
commited on
Commit
·
38e43b5
1
Parent(s):
3e0f8f8
Overall scores graph complete.
Browse files- .gitignore +2 -1
- app.py +361 -123
- img/council_icon.png +0 -0
- prompts.py +4 -1
.gitignore
CHANGED
@@ -1,2 +1,3 @@
|
|
1 |
env/
|
2 |
-
client_secret.json
|
|
|
|
1 |
env/
|
2 |
+
client_secret.json
|
3 |
+
__pycache__
|
app.py
CHANGED
@@ -15,10 +15,15 @@ from constants import (
|
|
15 |
LLM_TO_UI_NAME_MAP,
|
16 |
)
|
17 |
from prompts import *
|
18 |
-
from judging_dataclasses import
|
|
|
|
|
|
|
|
|
19 |
import pandas as pd
|
20 |
import seaborn as sns
|
21 |
import matplotlib.pyplot as plt
|
|
|
22 |
|
23 |
dotenv.load_dotenv()
|
24 |
|
@@ -67,6 +72,16 @@ def anthropic_streamlit_streamer(stream):
|
|
67 |
break # End of message, stop streaming
|
68 |
|
69 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
def google_streamlit_streamer(stream):
|
71 |
for chunk in stream:
|
72 |
yield chunk.text
|
@@ -146,22 +161,6 @@ def get_llm_response_stream(model_identifier, prompt):
|
|
146 |
return None
|
147 |
|
148 |
|
149 |
-
def get_response_key(model):
|
150 |
-
return model + "__response"
|
151 |
-
|
152 |
-
|
153 |
-
def get_model_from_response_key(response_key):
|
154 |
-
return response_key.split("__")[0]
|
155 |
-
|
156 |
-
|
157 |
-
def get_direct_assessment_judging_key(judge_model, response_model):
|
158 |
-
return "direct_assessment_judge__" + judge_model + "__" + response_model
|
159 |
-
|
160 |
-
|
161 |
-
def get_aggregator_response_key(model):
|
162 |
-
return model + "__aggregator_response"
|
163 |
-
|
164 |
-
|
165 |
def create_dataframe_for_direct_assessment_judging_response(
|
166 |
response: DirectAssessmentJudgingResponse,
|
167 |
):
|
@@ -203,21 +202,6 @@ def render_criteria_form(criteria_num):
|
|
203 |
)
|
204 |
|
205 |
|
206 |
-
def get_response_mapping():
|
207 |
-
# Inspect the session state for all the responses.
|
208 |
-
# This is a dictionary mapping model names to their responses.
|
209 |
-
# The aggregator response is also included in this mapping under the key "<model>__aggregator_response".
|
210 |
-
response_mapping = {}
|
211 |
-
for key in st.session_state.keys():
|
212 |
-
if "judge" in key:
|
213 |
-
continue
|
214 |
-
if key.endswith("__response"):
|
215 |
-
response_mapping[get_model_from_response_key(key)] = st.session_state[key]
|
216 |
-
if key.endswith("__aggregator_response"):
|
217 |
-
response_mapping[key] = st.session_state[key]
|
218 |
-
return response_mapping
|
219 |
-
|
220 |
-
|
221 |
def format_likert_comparison_options(options):
|
222 |
return "\n".join([f"{i + 1}: {option}" for i, option in enumerate(options)])
|
223 |
|
@@ -252,7 +236,7 @@ def get_default_direct_assessment_prompt(user_prompt):
|
|
252 |
def get_aggregator_prompt(aggregator_prompt, user_prompt, llms):
|
253 |
responses_from_other_llms = "\n\n".join(
|
254 |
[
|
255 |
-
f"{get_ui_friendly_name(model)} START\n{st.session_state
|
256 |
for model in llms
|
257 |
]
|
258 |
)
|
@@ -270,10 +254,6 @@ def get_default_aggregator_prompt(user_prompt, llms):
|
|
270 |
)
|
271 |
|
272 |
|
273 |
-
def get_ui_friendly_name(llm):
|
274 |
-
return LLM_TO_UI_NAME_MAP.get(llm, llm)
|
275 |
-
|
276 |
-
|
277 |
def get_parse_judging_response_for_direct_assessment_prompt(
|
278 |
judging_responses: dict[str, str],
|
279 |
criteria_list,
|
@@ -292,34 +272,58 @@ def get_parse_judging_response_for_direct_assessment_prompt(
|
|
292 |
)
|
293 |
|
294 |
|
295 |
-
|
296 |
-
|
297 |
-
|
298 |
-
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
|
306 |
-
|
307 |
-
|
308 |
-
|
309 |
-
|
310 |
-
|
311 |
-
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
|
321 |
-
|
322 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
323 |
|
324 |
|
325 |
def plot_criteria_scores(df):
|
@@ -364,6 +368,94 @@ def plot_criteria_scores(df):
|
|
364 |
st.pyplot(plt.gcf())
|
365 |
|
366 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
367 |
# Main Streamlit App
|
368 |
def main():
|
369 |
st.set_page_config(
|
@@ -395,7 +487,7 @@ def main():
|
|
395 |
|
396 |
# App title and description
|
397 |
st.title("Language Model Council Sandbox")
|
398 |
-
st.markdown("###### Invoke a council of LLMs to
|
399 |
st.markdown("###### [Paper](https://arxiv.org/abs/2406.08598)")
|
400 |
|
401 |
# Authentication system
|
@@ -413,18 +505,30 @@ def main():
|
|
413 |
st.error("Invalid credentials")
|
414 |
|
415 |
if st.session_state.authenticated:
|
416 |
-
|
|
|
417 |
|
418 |
# Council and aggregator selection
|
419 |
selected_models = llm_council_selector()
|
420 |
-
|
|
|
|
|
421 |
selected_aggregator = aggregator_selector()
|
422 |
|
|
|
|
|
|
|
|
|
|
|
|
|
423 |
# Prompt input
|
424 |
-
|
|
|
|
|
|
|
425 |
|
426 |
-
if
|
427 |
-
st.
|
428 |
|
429 |
response_columns = st.columns(3)
|
430 |
|
@@ -443,7 +547,7 @@ def main():
|
|
443 |
message_placeholder = st.empty()
|
444 |
stream = get_llm_response_stream(selected_model, user_prompt)
|
445 |
if stream:
|
446 |
-
st.session_state[
|
447 |
message_placeholder.write_stream(stream)
|
448 |
)
|
449 |
|
@@ -456,25 +560,25 @@ def main():
|
|
456 |
st.code(aggregator_prompt)
|
457 |
|
458 |
# Fetching and streaming response from the aggregator
|
459 |
-
st.write(
|
460 |
-
f"Mixture-of-Agents response from {get_ui_friendly_name(selected_aggregator)}"
|
461 |
-
)
|
462 |
with st.chat_message(
|
463 |
selected_aggregator,
|
464 |
-
avatar=
|
465 |
):
|
466 |
message_placeholder = st.empty()
|
467 |
aggregator_stream = get_llm_response_stream(
|
468 |
selected_aggregator, aggregator_prompt
|
469 |
)
|
470 |
if aggregator_stream:
|
471 |
-
|
472 |
-
|
473 |
-
|
474 |
-
|
|
|
|
|
475 |
|
476 |
# Judging.
|
477 |
-
st.markdown("#### Judging Configuration
|
478 |
|
479 |
# Choose the type of assessment
|
480 |
assessment_type = st.radio(
|
@@ -482,9 +586,48 @@ def main():
|
|
482 |
options=["Direct Assessment", "Pairwise Comparison"],
|
483 |
)
|
484 |
|
|
|
|
|
485 |
# Depending on the assessment type, render different forms
|
486 |
if assessment_type == "Direct Assessment":
|
487 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
488 |
direct_assessment_prompt = st.text_area(
|
489 |
"Prompt for the Direct Assessment",
|
490 |
value=get_default_direct_assessment_prompt(user_prompt=user_prompt),
|
@@ -495,10 +638,15 @@ def main():
|
|
495 |
criteria_list = DEFAULT_DIRECT_ASSESSMENT_CRITERIA_LIST
|
496 |
|
497 |
# Create DirectAssessment object when form is submitted
|
498 |
-
if
|
|
|
|
|
499 |
|
500 |
# Submit direct asssessment.
|
501 |
-
responses_for_judging =
|
|
|
|
|
|
|
502 |
|
503 |
response_judging_columns = st.columns(3)
|
504 |
|
@@ -515,11 +663,13 @@ def main():
|
|
515 |
]
|
516 |
|
517 |
with st_column:
|
518 |
-
if "
|
519 |
judging_model_header = "Mixture-of-Agents Response"
|
520 |
else:
|
521 |
judging_model_header = get_ui_friendly_name(response_model)
|
522 |
st.write(f"Judging for {judging_model_header}")
|
|
|
|
|
523 |
judging_prompt = get_direct_assessment_prompt(
|
524 |
direct_assessment_prompt=direct_assessment_prompt,
|
525 |
user_prompt=user_prompt,
|
@@ -543,18 +693,27 @@ def main():
|
|
543 |
judging_stream = get_llm_response_stream(
|
544 |
judging_model, judging_prompt
|
545 |
)
|
546 |
-
if judging_stream:
|
547 |
-
|
548 |
-
|
549 |
-
|
550 |
-
|
551 |
-
|
552 |
-
|
553 |
-
|
554 |
# When all of the judging is finished for the given response, get the actual
|
555 |
# values, parsed (use gpt-4o-mini for now) with json mode.
|
556 |
# TODO.
|
557 |
-
judging_responses =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
558 |
parse_judging_response_prompt = (
|
559 |
get_parse_judging_response_for_direct_assessment_prompt(
|
560 |
judging_responses,
|
@@ -562,45 +721,124 @@ def main():
|
|
562 |
SEVEN_POINT_DIRECT_ASSESSMENT_OPTIONS,
|
563 |
)
|
564 |
)
|
|
|
|
|
565 |
# Issue the prompt to openai mini with structured outputs
|
566 |
parsed_judging_responses = parse_judging_responses(
|
567 |
-
parse_judging_response_prompt
|
568 |
)
|
569 |
|
570 |
-
|
|
|
|
|
571 |
parsed_judging_responses
|
572 |
)
|
573 |
-
st.write(
|
|
|
|
|
|
|
|
|
574 |
|
575 |
-
|
576 |
-
|
577 |
-
|
578 |
-
|
|
|
579 |
|
580 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
581 |
|
582 |
-
|
583 |
-
|
584 |
-
|
585 |
-
|
586 |
-
|
587 |
-
|
588 |
-
|
589 |
-
|
590 |
-
|
591 |
-
|
592 |
-
|
593 |
-
|
594 |
-
|
595 |
-
|
596 |
-
|
597 |
-
|
598 |
-
|
599 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
600 |
)
|
601 |
-
|
602 |
-
#
|
603 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
604 |
|
605 |
else:
|
606 |
with cols[1]:
|
|
|
15 |
LLM_TO_UI_NAME_MAP,
|
16 |
)
|
17 |
from prompts import *
|
18 |
+
from judging_dataclasses import (
|
19 |
+
DirectAssessmentJudgingResponse,
|
20 |
+
DirectAssessmentCriterionScore,
|
21 |
+
DirectAssessmentCriteriaScores,
|
22 |
+
)
|
23 |
import pandas as pd
|
24 |
import seaborn as sns
|
25 |
import matplotlib.pyplot as plt
|
26 |
+
import numpy as np
|
27 |
|
28 |
dotenv.load_dotenv()
|
29 |
|
|
|
72 |
break # End of message, stop streaming
|
73 |
|
74 |
|
75 |
+
def get_ui_friendly_name(llm):
|
76 |
+
if "agg__" in llm:
|
77 |
+
return (
|
78 |
+
"MoA ("
|
79 |
+
+ LLM_TO_UI_NAME_MAP.get(llm.split("__")[1], llm.split("__")[1])
|
80 |
+
+ ")"
|
81 |
+
)
|
82 |
+
return LLM_TO_UI_NAME_MAP.get(llm, llm)
|
83 |
+
|
84 |
+
|
85 |
def google_streamlit_streamer(stream):
|
86 |
for chunk in stream:
|
87 |
yield chunk.text
|
|
|
161 |
return None
|
162 |
|
163 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
164 |
def create_dataframe_for_direct_assessment_judging_response(
|
165 |
response: DirectAssessmentJudgingResponse,
|
166 |
):
|
|
|
202 |
)
|
203 |
|
204 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
205 |
def format_likert_comparison_options(options):
|
206 |
return "\n".join([f"{i + 1}: {option}" for i, option in enumerate(options)])
|
207 |
|
|
|
236 |
def get_aggregator_prompt(aggregator_prompt, user_prompt, llms):
|
237 |
responses_from_other_llms = "\n\n".join(
|
238 |
[
|
239 |
+
f"{get_ui_friendly_name(model)} START\n{st.session_state['responses'][model]}\n\n{get_ui_friendly_name(model)} END\n\n\n"
|
240 |
for model in llms
|
241 |
]
|
242 |
)
|
|
|
254 |
)
|
255 |
|
256 |
|
|
|
|
|
|
|
|
|
257 |
def get_parse_judging_response_for_direct_assessment_prompt(
|
258 |
judging_responses: dict[str, str],
|
259 |
criteria_list,
|
|
|
272 |
)
|
273 |
|
274 |
|
275 |
+
DEBUG_MODE = True
|
276 |
+
|
277 |
+
|
278 |
+
def parse_judging_responses(
|
279 |
+
prompt: str, judging_responses: dict[str, str]
|
280 |
+
) -> DirectAssessmentJudgingResponse:
|
281 |
+
if DEBUG_MODE:
|
282 |
+
return DirectAssessmentJudgingResponse(
|
283 |
+
judging_models=[
|
284 |
+
DirectAssessmentCriteriaScores(
|
285 |
+
model="together://meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
|
286 |
+
criteria_scores=[
|
287 |
+
DirectAssessmentCriterionScore(
|
288 |
+
criterion="helpfulness", score=3, explanation="explanation1"
|
289 |
+
),
|
290 |
+
DirectAssessmentCriterionScore(
|
291 |
+
criterion="conciseness", score=4, explanation="explanation2"
|
292 |
+
),
|
293 |
+
DirectAssessmentCriterionScore(
|
294 |
+
criterion="relevance", score=5, explanation="explanation3"
|
295 |
+
),
|
296 |
+
],
|
297 |
+
),
|
298 |
+
DirectAssessmentCriteriaScores(
|
299 |
+
model="together://meta-llama/Llama-3.2-3B-Instruct-Turbo",
|
300 |
+
criteria_scores=[
|
301 |
+
DirectAssessmentCriterionScore(
|
302 |
+
criterion="helpfulness", score=1, explanation="explanation1"
|
303 |
+
),
|
304 |
+
DirectAssessmentCriterionScore(
|
305 |
+
criterion="conciseness", score=2, explanation="explanation2"
|
306 |
+
),
|
307 |
+
DirectAssessmentCriterionScore(
|
308 |
+
criterion="relevance", score=3, explanation="explanation3"
|
309 |
+
),
|
310 |
+
],
|
311 |
+
),
|
312 |
+
]
|
313 |
+
)
|
314 |
+
else:
|
315 |
+
completion = client.beta.chat.completions.parse(
|
316 |
+
model="gpt-4o-mini",
|
317 |
+
messages=[
|
318 |
+
{
|
319 |
+
"role": "system",
|
320 |
+
"content": "Parse the judging responses into structured data.",
|
321 |
+
},
|
322 |
+
{"role": "user", "content": prompt},
|
323 |
+
],
|
324 |
+
response_format=DirectAssessmentJudgingResponse,
|
325 |
+
)
|
326 |
+
return completion.choices[0].message.parsed
|
327 |
|
328 |
|
329 |
def plot_criteria_scores(df):
|
|
|
368 |
st.pyplot(plt.gcf())
|
369 |
|
370 |
|
371 |
+
def plot_overall_scores(overall_scores_df):
|
372 |
+
# Calculate mean and standard deviation
|
373 |
+
summary = (
|
374 |
+
overall_scores_df.groupby("response_model")
|
375 |
+
.agg({"score": ["mean", "std"]})
|
376 |
+
.reset_index()
|
377 |
+
)
|
378 |
+
summary.columns = ["response_model", "mean_score", "std_score"]
|
379 |
+
|
380 |
+
# Add UI-friendly names
|
381 |
+
summary["ui_friendly_name"] = summary["response_model"].apply(get_ui_friendly_name)
|
382 |
+
|
383 |
+
# Sort the summary dataframe by mean_score in descending order
|
384 |
+
summary = summary.sort_values("mean_score", ascending=False)
|
385 |
+
|
386 |
+
# Create the plot
|
387 |
+
plt.figure(figsize=(8, 5))
|
388 |
+
|
389 |
+
# Plot bars with rainbow colors
|
390 |
+
ax = sns.barplot(
|
391 |
+
x="ui_friendly_name",
|
392 |
+
y="mean_score",
|
393 |
+
data=summary,
|
394 |
+
palette="prism",
|
395 |
+
capsize=0.1,
|
396 |
+
)
|
397 |
+
|
398 |
+
# Add error bars manually
|
399 |
+
x_coords = range(len(summary))
|
400 |
+
plt.errorbar(
|
401 |
+
x=x_coords,
|
402 |
+
y=summary["mean_score"],
|
403 |
+
yerr=summary["std_score"],
|
404 |
+
fmt="none",
|
405 |
+
c="black",
|
406 |
+
capsize=5,
|
407 |
+
zorder=10, # Ensure error bars are on top
|
408 |
+
)
|
409 |
+
|
410 |
+
# Add text annotations
|
411 |
+
for i, row in summary.iterrows():
|
412 |
+
ax.text(
|
413 |
+
i,
|
414 |
+
row["mean_score"],
|
415 |
+
f"{row['mean_score']:.2f}",
|
416 |
+
ha="center",
|
417 |
+
va="bottom",
|
418 |
+
fontweight="bold",
|
419 |
+
color="black",
|
420 |
+
bbox=dict(facecolor="white", edgecolor="none", alpha=0.7, pad=0.5),
|
421 |
+
)
|
422 |
+
|
423 |
+
# Customize the plot
|
424 |
+
plt.xlabel("")
|
425 |
+
plt.ylabel("Overall Score")
|
426 |
+
plt.xticks(rotation=45, ha="right")
|
427 |
+
plt.tight_layout()
|
428 |
+
|
429 |
+
# Display the plot in Streamlit
|
430 |
+
st.pyplot(plt.gcf())
|
431 |
+
|
432 |
+
|
433 |
+
def plot_per_judge_overall_scores(df):
|
434 |
+
# Find the overall score by finding the overall score for each judge, and then averaging
|
435 |
+
# over all judges.
|
436 |
+
grouped = df.groupby(["llm_judge_model"]).agg({"score": ["mean"]}).reset_index()
|
437 |
+
grouped.columns = ["llm_judge_model", "overall_score"]
|
438 |
+
|
439 |
+
# Create the horizontal bar plot
|
440 |
+
plt.figure(figsize=(10, 6))
|
441 |
+
ax = sns.barplot(
|
442 |
+
data=grouped,
|
443 |
+
y="llm_judge_model",
|
444 |
+
x="overall_score",
|
445 |
+
hue="llm_judge_model",
|
446 |
+
orient="h",
|
447 |
+
)
|
448 |
+
|
449 |
+
# Customize the plot
|
450 |
+
plt.title("Overall Scores by LLM Judge Model")
|
451 |
+
plt.xlabel("Overall Score")
|
452 |
+
plt.ylabel("LLM Judge Model")
|
453 |
+
|
454 |
+
# Adjust layout and display the plot
|
455 |
+
plt.tight_layout()
|
456 |
+
st.pyplot(plt)
|
457 |
+
|
458 |
+
|
459 |
# Main Streamlit App
|
460 |
def main():
|
461 |
st.set_page_config(
|
|
|
487 |
|
488 |
# App title and description
|
489 |
st.title("Language Model Council Sandbox")
|
490 |
+
st.markdown("###### Invoke a council of LLMs to judge each other's responses.")
|
491 |
st.markdown("###### [Paper](https://arxiv.org/abs/2406.08598)")
|
492 |
|
493 |
# Authentication system
|
|
|
505 |
st.error("Invalid credentials")
|
506 |
|
507 |
if st.session_state.authenticated:
|
508 |
+
# cols[1].success("Logged in successfully!")
|
509 |
+
st.markdown("#### LLM Council Member Selection")
|
510 |
|
511 |
# Council and aggregator selection
|
512 |
selected_models = llm_council_selector()
|
513 |
+
|
514 |
+
# st.write("Selected Models:", selected_models)
|
515 |
+
|
516 |
selected_aggregator = aggregator_selector()
|
517 |
|
518 |
+
# Initialize session state for collecting responses.
|
519 |
+
if "responses" not in st.session_state:
|
520 |
+
st.session_state.responses = {}
|
521 |
+
# if "aggregator_response" not in st.session_state:
|
522 |
+
# st.session_state.aggregator_response = {}
|
523 |
+
|
524 |
# Prompt input
|
525 |
+
st.markdown("#### Enter your prompt")
|
526 |
+
_, center_column, _ = st.columns([3, 5, 3])
|
527 |
+
with center_column:
|
528 |
+
user_prompt = st.text_area(value="Say 'Hello World'", label="")
|
529 |
|
530 |
+
if center_column.button("Submit", use_container_width=True):
|
531 |
+
st.markdown("#### Responses")
|
532 |
|
533 |
response_columns = st.columns(3)
|
534 |
|
|
|
547 |
message_placeholder = st.empty()
|
548 |
stream = get_llm_response_stream(selected_model, user_prompt)
|
549 |
if stream:
|
550 |
+
st.session_state["responses"][selected_model] = (
|
551 |
message_placeholder.write_stream(stream)
|
552 |
)
|
553 |
|
|
|
560 |
st.code(aggregator_prompt)
|
561 |
|
562 |
# Fetching and streaming response from the aggregator
|
563 |
+
st.write(f"Mixture-of-Agents ({get_ui_friendly_name(selected_aggregator)})")
|
|
|
|
|
564 |
with st.chat_message(
|
565 |
selected_aggregator,
|
566 |
+
avatar="img/council_icon.png",
|
567 |
):
|
568 |
message_placeholder = st.empty()
|
569 |
aggregator_stream = get_llm_response_stream(
|
570 |
selected_aggregator, aggregator_prompt
|
571 |
)
|
572 |
if aggregator_stream:
|
573 |
+
st.session_state["responses"]["agg__" + selected_aggregator] = (
|
574 |
+
message_placeholder.write_stream(aggregator_stream)
|
575 |
+
)
|
576 |
+
|
577 |
+
# st.write("Responses (in session state):")
|
578 |
+
# st.write(st.session_state["responses"])
|
579 |
|
580 |
# Judging.
|
581 |
+
st.markdown("#### Judging Configuration")
|
582 |
|
583 |
# Choose the type of assessment
|
584 |
assessment_type = st.radio(
|
|
|
586 |
options=["Direct Assessment", "Pairwise Comparison"],
|
587 |
)
|
588 |
|
589 |
+
_, center_column, _ = st.columns([3, 5, 3])
|
590 |
+
|
591 |
# Depending on the assessment type, render different forms
|
592 |
if assessment_type == "Direct Assessment":
|
593 |
+
|
594 |
+
# Initialize session state for direct assessment.
|
595 |
+
if "direct_assessment_overall_score" not in st.session_state:
|
596 |
+
st.session_state["direct_assessment_overall_score"] = {}
|
597 |
+
if "direct_assessment_judging_df" not in st.session_state:
|
598 |
+
st.session_state["direct_assessment_judging_df"] = {}
|
599 |
+
for response_model in selected_models:
|
600 |
+
st.session_state["direct_assessment_judging_df"][
|
601 |
+
response_model
|
602 |
+
] = {}
|
603 |
+
# aggregator model
|
604 |
+
st.session_state["direct_assessment_judging_df"][
|
605 |
+
"agg__" + selected_aggregator
|
606 |
+
] = {}
|
607 |
+
if "direct_assessment_judging_responses" not in st.session_state:
|
608 |
+
st.session_state["direct_assessment_judging_responses"] = {}
|
609 |
+
for response_model in selected_models:
|
610 |
+
st.session_state["direct_assessment_judging_responses"][
|
611 |
+
response_model
|
612 |
+
] = {}
|
613 |
+
# aggregator model
|
614 |
+
st.session_state["direct_assessment_judging_responses"][
|
615 |
+
"agg__" + selected_aggregator
|
616 |
+
] = {}
|
617 |
+
if "direct_assessment_overall_scores" not in st.session_state:
|
618 |
+
st.session_state["direct_assessment_overall_scores"] = {}
|
619 |
+
for response_model in selected_models:
|
620 |
+
st.session_state["direct_assessment_overall_scores"][
|
621 |
+
response_model
|
622 |
+
] = {}
|
623 |
+
st.session_state["direct_assessment_overall_scores"][
|
624 |
+
"agg__" + selected_aggregator
|
625 |
+
] = {}
|
626 |
+
if "judging_status" not in st.session_state:
|
627 |
+
st.session_state["judging_status"] = "incomplete"
|
628 |
+
|
629 |
+
# Direct assessment prompt.
|
630 |
+
with center_column.expander("Direct Assessment Prompt"):
|
631 |
direct_assessment_prompt = st.text_area(
|
632 |
"Prompt for the Direct Assessment",
|
633 |
value=get_default_direct_assessment_prompt(user_prompt=user_prompt),
|
|
|
638 |
criteria_list = DEFAULT_DIRECT_ASSESSMENT_CRITERIA_LIST
|
639 |
|
640 |
# Create DirectAssessment object when form is submitted
|
641 |
+
if center_column.button(
|
642 |
+
"Submit Direct Assessment", use_container_width=True
|
643 |
+
):
|
644 |
|
645 |
# Submit direct asssessment.
|
646 |
+
responses_for_judging = st.session_state["responses"]
|
647 |
+
|
648 |
+
# st.write("Responses for judging (in session state):")
|
649 |
+
# st.write(responses_for_judging)
|
650 |
|
651 |
response_judging_columns = st.columns(3)
|
652 |
|
|
|
663 |
]
|
664 |
|
665 |
with st_column:
|
666 |
+
if "agg__" in response_model:
|
667 |
judging_model_header = "Mixture-of-Agents Response"
|
668 |
else:
|
669 |
judging_model_header = get_ui_friendly_name(response_model)
|
670 |
st.write(f"Judging for {judging_model_header}")
|
671 |
+
# st.write("Response being judged: ")
|
672 |
+
# st.write(response)
|
673 |
judging_prompt = get_direct_assessment_prompt(
|
674 |
direct_assessment_prompt=direct_assessment_prompt,
|
675 |
user_prompt=user_prompt,
|
|
|
693 |
judging_stream = get_llm_response_stream(
|
694 |
judging_model, judging_prompt
|
695 |
)
|
696 |
+
# if judging_stream:
|
697 |
+
st.session_state[
|
698 |
+
"direct_assessment_judging_responses"
|
699 |
+
][response_model][
|
700 |
+
judging_model
|
701 |
+
] = message_placeholder.write_stream(
|
702 |
+
judging_stream
|
703 |
+
)
|
704 |
# When all of the judging is finished for the given response, get the actual
|
705 |
# values, parsed (use gpt-4o-mini for now) with json mode.
|
706 |
# TODO.
|
707 |
+
judging_responses = st.session_state[
|
708 |
+
"direct_assessment_judging_responses"
|
709 |
+
][response_model]
|
710 |
+
|
711 |
+
# st.write("Judging responses (in session state):")
|
712 |
+
# st.write(judging_responses)
|
713 |
+
|
714 |
+
if not judging_responses:
|
715 |
+
st.error(f"No judging responses for {response_model}")
|
716 |
+
quit()
|
717 |
parse_judging_response_prompt = (
|
718 |
get_parse_judging_response_for_direct_assessment_prompt(
|
719 |
judging_responses,
|
|
|
721 |
SEVEN_POINT_DIRECT_ASSESSMENT_OPTIONS,
|
722 |
)
|
723 |
)
|
724 |
+
with st.expander("Parse Judging Response Prompt"):
|
725 |
+
st.code(parse_judging_response_prompt)
|
726 |
# Issue the prompt to openai mini with structured outputs
|
727 |
parsed_judging_responses = parse_judging_responses(
|
728 |
+
parse_judging_response_prompt, judging_responses
|
729 |
)
|
730 |
|
731 |
+
st.session_state["direct_assessment_judging_df"][
|
732 |
+
response_model
|
733 |
+
] = create_dataframe_for_direct_assessment_judging_response(
|
734 |
parsed_judging_responses
|
735 |
)
|
736 |
+
st.write(
|
737 |
+
st.session_state["direct_assessment_judging_df"][
|
738 |
+
response_model
|
739 |
+
]
|
740 |
+
)
|
741 |
|
742 |
+
plot_criteria_scores(
|
743 |
+
st.session_state["direct_assessment_judging_df"][
|
744 |
+
response_model
|
745 |
+
]
|
746 |
+
)
|
747 |
|
748 |
+
# Find the overall score by finding the overall score for each judge, and then averaging
|
749 |
+
# over all judges.
|
750 |
+
plot_per_judge_overall_scores(
|
751 |
+
st.session_state["direct_assessment_judging_df"][
|
752 |
+
response_model
|
753 |
+
]
|
754 |
+
)
|
755 |
|
756 |
+
grouped = (
|
757 |
+
st.session_state["direct_assessment_judging_df"][
|
758 |
+
response_model
|
759 |
+
]
|
760 |
+
.groupby(["llm_judge_model"])
|
761 |
+
.agg({"score": ["mean"]})
|
762 |
+
.reset_index()
|
763 |
+
)
|
764 |
+
grouped.columns = ["llm_judge_model", "overall_score"]
|
765 |
+
|
766 |
+
# st.write(
|
767 |
+
# "Extracting overall scores from this grouped dataframe:"
|
768 |
+
# )
|
769 |
+
# st.write(grouped)
|
770 |
+
|
771 |
+
# Save the overall scores to the session state.
|
772 |
+
for record in grouped.to_dict(orient="records"):
|
773 |
+
st.session_state["direct_assessment_overall_scores"][
|
774 |
+
response_model
|
775 |
+
][record["llm_judge_model"]] = record["overall_score"]
|
776 |
+
|
777 |
+
overall_score = grouped["overall_score"].mean()
|
778 |
+
controversy = grouped["overall_score"].std()
|
779 |
+
st.write(f"Overall Score: {overall_score:.2f}")
|
780 |
+
st.write(f"Controversy: {controversy:.2f}")
|
781 |
+
|
782 |
+
st.session_state["judging_status"] = "complete"
|
783 |
+
|
784 |
+
# Judging is complete.
|
785 |
+
st.write("#### Results")
|
786 |
+
# The session state now contains the overall scores for each response from each judge.
|
787 |
+
if st.session_state["judging_status"] == "complete":
|
788 |
+
overall_scores_df_raw = pd.DataFrame(
|
789 |
+
st.session_state["direct_assessment_overall_scores"]
|
790 |
+
).reset_index()
|
791 |
+
|
792 |
+
overall_scores_df = pd.melt(
|
793 |
+
overall_scores_df_raw,
|
794 |
+
id_vars=["index"],
|
795 |
+
var_name="response_model",
|
796 |
+
value_name="score",
|
797 |
+
).rename(columns={"index": "judging_model"})
|
798 |
+
|
799 |
+
# Print the overall winner.
|
800 |
+
overall_winner = overall_scores_df.loc[
|
801 |
+
overall_scores_df["score"].idxmax()
|
802 |
+
]
|
803 |
+
|
804 |
+
st.write(
|
805 |
+
f"**Overall Winner:** {get_ui_friendly_name(overall_winner['response_model'])}"
|
806 |
)
|
807 |
+
# Find how much the standard deviation overlaps with other models.
|
808 |
+
# Calculate separability.
|
809 |
+
# TODO.
|
810 |
+
st.write(f"**Confidence:** {overall_winner['score']:.2f}")
|
811 |
+
|
812 |
+
left_column, right_column = st.columns([1, 1])
|
813 |
+
with left_column:
|
814 |
+
plot_overall_scores(overall_scores_df)
|
815 |
+
|
816 |
+
with right_column:
|
817 |
+
st.dataframe(overall_scores_df)
|
818 |
+
|
819 |
+
elif assessment_type == "Pairwise Comparison":
|
820 |
+
pass
|
821 |
+
# pairwise_comparison_prompt = st.text_area(
|
822 |
+
# "Prompt for the Pairwise Comparison"
|
823 |
+
# )
|
824 |
+
# granularity = st.selectbox("Granularity", ["coarse", "fine", "super fine"])
|
825 |
+
# ties_allowed = st.checkbox("Are ties allowed?")
|
826 |
+
# position_swapping = st.checkbox("Enable position swapping?")
|
827 |
+
# reference_model = st.text_input("Reference Model")
|
828 |
+
|
829 |
+
# # Create PairwiseComparison object when form is submitted
|
830 |
+
# if st.button("Submit Pairwise Comparison"):
|
831 |
+
# pairwise_comparison_config = PairwiseComparison(
|
832 |
+
# type="pairwise_comparison",
|
833 |
+
# granularity=granularity,
|
834 |
+
# ties_allowed=ties_allowed,
|
835 |
+
# position_swapping=position_swapping,
|
836 |
+
# reference_model=reference_model,
|
837 |
+
# prompt=prompt,
|
838 |
+
# )
|
839 |
+
# st.success(f"Pairwise Comparison Created: {pairwise_comparison_config}")
|
840 |
+
# # Submit pairwise comparison.
|
841 |
+
# responses_for_judging = st.session_state["responses"]
|
842 |
|
843 |
else:
|
844 |
with cols[1]:
|
img/council_icon.png
ADDED
prompts.py
CHANGED
@@ -25,7 +25,10 @@ DEFAULT_AGGREGATOR_PROMPT = """We are trying to come up with the best response t
|
|
25 |
Responses from other LLMs:
|
26 |
{responses_from_other_llms}
|
27 |
|
28 |
-
|
|
|
|
|
|
|
29 |
|
30 |
|
31 |
DEFAULT_DIRECT_ASSESSMENT_PROMPT = """We are trying to assess the quality of a response to a user query.
|
|
|
25 |
Responses from other LLMs:
|
26 |
{responses_from_other_llms}
|
27 |
|
28 |
+
Consider how you would combine the best aspects of the responses above into a single response.
|
29 |
+
|
30 |
+
Directly provide your response to the user's query as if you were the original LLM. Do not mention that you are synthesizing the responses from other LLMs.
|
31 |
+
"""
|
32 |
|
33 |
|
34 |
DEFAULT_DIRECT_ASSESSMENT_PROMPT = """We are trying to assess the quality of a response to a user query.
|