Spaces:
Running
Running
clean up code
Browse files- app.py +1 -2
- eval_modules/calc_repetitions_v2e.py +0 -1333
- eval_modules/utils.py +67 -156
app.py
CHANGED
@@ -17,8 +17,7 @@ path = os.path.dirname(found_dotenv)
|
|
17 |
print(f"Adding {path} to sys.path")
|
18 |
sys.path.append(path)
|
19 |
|
20 |
-
from eval_modules.utils import calc_perf_scores
|
21 |
-
from eval_modules.calc_repetitions_v2e import detect_repetitions
|
22 |
|
23 |
model_name = os.getenv("MODEL_NAME") or "microsoft/Phi-3.5-mini-instruct"
|
24 |
hf_token = os.getenv("HF_TOKEN")
|
|
|
17 |
print(f"Adding {path} to sys.path")
|
18 |
sys.path.append(path)
|
19 |
|
20 |
+
from eval_modules.utils import calc_perf_scores, detect_repetitions
|
|
|
21 |
|
22 |
model_name = os.getenv("MODEL_NAME") or "microsoft/Phi-3.5-mini-instruct"
|
23 |
hf_token = os.getenv("HF_TOKEN")
|
eval_modules/calc_repetitions_v2e.py
DELETED
@@ -1,1333 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
import re
|
3 |
-
import math
|
4 |
-
import pandas as pd
|
5 |
-
import numpy as np
|
6 |
-
import matplotlib.pyplot as plt
|
7 |
-
import matplotlib.ticker as mtick
|
8 |
-
import seaborn as sns
|
9 |
-
import nltk
|
10 |
-
import evaluate
|
11 |
-
import traceback
|
12 |
-
|
13 |
-
bert_score = evaluate.load("bertscore")
|
14 |
-
meteor = evaluate.load("meteor")
|
15 |
-
|
16 |
-
print(f"loading: {__file__}")
|
17 |
-
|
18 |
-
# pattern_non_word_char_repetition = re.compile(r"\s{5,}")
|
19 |
-
# pattern_text_repetitions = re.compile(r"(.{5}.*)\s*((\1)\s*)+", re.M | re.DOTALL)
|
20 |
-
|
21 |
-
# final version
|
22 |
-
pattern_non_word_char_repetition = re.compile(r"[\s\W]{5,}")
|
23 |
-
pattern_text_repetitions = re.compile(
|
24 |
-
r"(?P<repeat>.{5}.*?)(?:[\s\W]*(?P=repeat))+", re.M | re.DOTALL | re.IGNORECASE
|
25 |
-
)
|
26 |
-
# Explanation of the Regex Pattern:
|
27 |
-
# (?P<repeat>.{5}.*?): Captures any sequence of characters with minimal length of 5 and names this group repeat.
|
28 |
-
# .*?: Matches zero or more characters, non-greedily (as few as possible).
|
29 |
-
# (?:[\s\W]+(?P=repeat))+: A non-capturing group that matches one or more repetitions of:
|
30 |
-
# [\s\W]+: One or more whitespace or non-word characters (spaces, punctuation, etc.).
|
31 |
-
# (?P=repeat): A backreference to the named group repeat.
|
32 |
-
|
33 |
-
|
34 |
-
def del_non_word_char_repetition(text, debug=False):
|
35 |
-
count = 0
|
36 |
-
|
37 |
-
if isinstance(text, str):
|
38 |
-
if debug:
|
39 |
-
print("----detect non-word characters repetition----")
|
40 |
-
count = len(text)
|
41 |
-
text = pattern_non_word_char_repetition.sub("\t", text)
|
42 |
-
count -= len(text)
|
43 |
-
if debug and count:
|
44 |
-
print(f"removed non-word characters repetition: {count}")
|
45 |
-
return text, count
|
46 |
-
|
47 |
-
|
48 |
-
# final version for repetition detection
|
49 |
-
def detect_text_repetitions(text, debug=False):
|
50 |
-
count = 0
|
51 |
-
|
52 |
-
if isinstance(text, str):
|
53 |
-
if debug:
|
54 |
-
print("----detect text repetitions----")
|
55 |
-
matches = pattern_text_repetitions.finditer(text)
|
56 |
-
for match in matches:
|
57 |
-
if debug:
|
58 |
-
print(match)
|
59 |
-
for groupNum in range(0, len(match.groups())):
|
60 |
-
groupNum = groupNum + 1
|
61 |
-
print(
|
62 |
-
"Group {groupNum} found at {start}-{end}: `{group}`".format(
|
63 |
-
groupNum=groupNum,
|
64 |
-
start=match.start(groupNum),
|
65 |
-
end=match.end(groupNum),
|
66 |
-
group=match.group(groupNum),
|
67 |
-
)
|
68 |
-
)
|
69 |
-
|
70 |
-
start, end = match.span()
|
71 |
-
count += end - start - len(match.group(1))
|
72 |
-
|
73 |
-
return count
|
74 |
-
|
75 |
-
|
76 |
-
def detect_repetitions(text, debug=False):
|
77 |
-
if isinstance(text, str) is False:
|
78 |
-
return 0, 0, 0
|
79 |
-
text, count_non_word_char_repetition = del_non_word_char_repetition(
|
80 |
-
text, debug=debug
|
81 |
-
)
|
82 |
-
count_text_repetitions = detect_text_repetitions(text, debug=debug)
|
83 |
-
total_repetitions = count_non_word_char_repetition + count_text_repetitions
|
84 |
-
|
85 |
-
result = (count_non_word_char_repetition, count_text_repetitions, total_repetitions)
|
86 |
-
|
87 |
-
if debug:
|
88 |
-
print(result)
|
89 |
-
return result
|
90 |
-
|
91 |
-
|
92 |
-
def detect_scores(
|
93 |
-
row, debug=False, answer_col="answer", ground_truth_col="ground_truth"
|
94 |
-
):
|
95 |
-
newline_score, repetition_score, total_repetitions = detect_repetitions(
|
96 |
-
row[answer_col], debug=debug
|
97 |
-
)
|
98 |
-
|
99 |
-
if ground_truth_col:
|
100 |
-
ground_truth_newline_score, ground_truth_repetition_score, _ = (
|
101 |
-
detect_repetitions(row[ground_truth_col], debug=debug)
|
102 |
-
)
|
103 |
-
|
104 |
-
newline_score -= ground_truth_newline_score
|
105 |
-
if newline_score < 0:
|
106 |
-
newline_score = 0
|
107 |
-
|
108 |
-
repetition_score -= ground_truth_repetition_score
|
109 |
-
if repetition_score < 0:
|
110 |
-
repetition_score = 0
|
111 |
-
|
112 |
-
total_repetitions = newline_score + repetition_score
|
113 |
-
|
114 |
-
return pd.Series([newline_score, repetition_score, total_repetitions])
|
115 |
-
|
116 |
-
|
117 |
-
def load_with_newline_and_repetition_scores(result_file, force_recalculate=False):
|
118 |
-
print(f"loading result file: {result_file}")
|
119 |
-
df = pd.read_csv(result_file, comment="#", on_bad_lines="warn")
|
120 |
-
|
121 |
-
if (
|
122 |
-
force_recalculate
|
123 |
-
or "newline_score" not in df.columns
|
124 |
-
or "repetition_score" not in df.columns
|
125 |
-
or "total_repetitions" not in df.columns
|
126 |
-
or "nrr" not in df.columns
|
127 |
-
or "rr" not in df.columns
|
128 |
-
):
|
129 |
-
if (
|
130 |
-
force_recalculate
|
131 |
-
or "newline_score" not in df.columns
|
132 |
-
or "repetition_score" not in df.columns
|
133 |
-
or "total_repetitions" not in df.columns
|
134 |
-
):
|
135 |
-
df[["newline_score", "repetition_score", "total_repetitions"]] = df.apply(
|
136 |
-
detect_scores, axis=1
|
137 |
-
)
|
138 |
-
|
139 |
-
df["answer_len"] = df["answer"].apply(
|
140 |
-
lambda x: len(x) if isinstance(x, str) else 0
|
141 |
-
)
|
142 |
-
|
143 |
-
df["nrr"] = df.apply(
|
144 |
-
lambda x: (
|
145 |
-
1
|
146 |
-
if x["answer_len"] == 0
|
147 |
-
else 1 - (x["newline_score"] + x["repetition_score"]) / x["answer_len"]
|
148 |
-
),
|
149 |
-
axis=1,
|
150 |
-
)
|
151 |
-
|
152 |
-
df["rr"] = df["nrr"].apply(lambda x: 1 - x)
|
153 |
-
|
154 |
-
df.to_csv(result_file, index=False)
|
155 |
-
|
156 |
-
return df
|
157 |
-
|
158 |
-
|
159 |
-
def replace_last(source_string, old_string, new_string):
|
160 |
-
head, _sep, tail = source_string.rpartition(old_string)
|
161 |
-
return head + new_string + tail
|
162 |
-
|
163 |
-
|
164 |
-
def load_for_repetition_penalty(
|
165 |
-
csv_result_file, repetition_penalty, force_recalculate=False
|
166 |
-
):
|
167 |
-
result_file = replace_last(
|
168 |
-
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}.csv"
|
169 |
-
)
|
170 |
-
return load_with_newline_and_repetition_scores(
|
171 |
-
result_file, force_recalculate=force_recalculate
|
172 |
-
)
|
173 |
-
|
174 |
-
|
175 |
-
rap_penalty_functions = {
|
176 |
-
"linear": lambda x: x,
|
177 |
-
"quadratic": lambda x: x * x,
|
178 |
-
"cubic": lambda x: x * x * x,
|
179 |
-
"logarithmic": lambda x: math.log(x + 1, 2),
|
180 |
-
"exponential": lambda x: math.exp(x - 1),
|
181 |
-
}
|
182 |
-
|
183 |
-
|
184 |
-
def calc_adjusted_performance(f, r, l=1, penalty_function="cubic"):
|
185 |
-
n = 1 - r / l if l > 0 else 0
|
186 |
-
return f * rap_penalty_functions[penalty_function](n)
|
187 |
-
|
188 |
-
|
189 |
-
def calculate_adjusted_performance(row):
|
190 |
-
r = row["total_repetitions"]
|
191 |
-
l = row["answer_len"]
|
192 |
-
adjusted_precision = calc_adjusted_performance(row["precision"], r, l)
|
193 |
-
adjusted_recall = calc_adjusted_performance(row["recall"], r, l)
|
194 |
-
return pd.Series([adjusted_precision, adjusted_recall])
|
195 |
-
|
196 |
-
|
197 |
-
def load_performance_df(csv_result_file, repetition_penalty):
|
198 |
-
result_file = replace_last(
|
199 |
-
csv_result_file, ".csv", f"_RP_{repetition_penalty:.3f}-t2_evaluated.json"
|
200 |
-
)
|
201 |
-
result_file = result_file.replace("/results/", "/eval/")
|
202 |
-
print(f"loading json file: {result_file}")
|
203 |
-
df = pd.read_json(result_file)
|
204 |
-
|
205 |
-
return df
|
206 |
-
|
207 |
-
|
208 |
-
def calculate_performance_score(
|
209 |
-
csv_result_file, repetition_penalty, force_recalculate=False
|
210 |
-
):
|
211 |
-
result_file = replace_last(
|
212 |
-
csv_result_file, ".csv", f"_rpp_{repetition_penalty:.2f}.csv"
|
213 |
-
)
|
214 |
-
|
215 |
-
if os.path.exists(result_file):
|
216 |
-
print(f"loading result file: {result_file}")
|
217 |
-
df = load_with_newline_and_repetition_scores(
|
218 |
-
result_file, force_recalculate=force_recalculate
|
219 |
-
)
|
220 |
-
else:
|
221 |
-
print(f"re-creating result file: {result_file}")
|
222 |
-
df = pd.DataFrame()
|
223 |
-
force_recalculate = True
|
224 |
-
|
225 |
-
if force_recalculate or "f2" in df.columns or "f1" not in df.columns:
|
226 |
-
try:
|
227 |
-
perf_df = load_performance_df(csv_result_file, repetition_penalty)
|
228 |
-
df.drop(
|
229 |
-
columns=[
|
230 |
-
"precision",
|
231 |
-
"recall",
|
232 |
-
"f1",
|
233 |
-
"f2",
|
234 |
-
"entities_in_answer",
|
235 |
-
"entities_in_question",
|
236 |
-
"word_count",
|
237 |
-
],
|
238 |
-
errors="ignore",
|
239 |
-
inplace=True,
|
240 |
-
)
|
241 |
-
|
242 |
-
df["id"] = perf_df["id"]
|
243 |
-
df["question"] = perf_df["question"]
|
244 |
-
df["answer"] = perf_df["pred_answer"]
|
245 |
-
df["word_count"] = df["answer"].apply(
|
246 |
-
lambda x: len(nltk.word_tokenize(x)) if isinstance(x, str) else 0
|
247 |
-
)
|
248 |
-
df["ground_truth"] = perf_df["ground_truth"]
|
249 |
-
|
250 |
-
df["eval_gemini_1.0_pro"] = perf_df["eval_gemini_1.0_pro"]
|
251 |
-
df["precision"] = perf_df["score"].apply(lambda x: x[0])
|
252 |
-
df["recall"] = perf_df["score"].apply(lambda x: x[1])
|
253 |
-
df["f1"] = perf_df["score"].apply(lambda x: x[2])
|
254 |
-
except Exception as e:
|
255 |
-
print(f"\tignored error: {e}")
|
256 |
-
# traceback.print_exc()
|
257 |
-
|
258 |
-
df[["newline_score", "repetition_score", "total_repetitions"]] = df.apply(
|
259 |
-
detect_scores, axis=1
|
260 |
-
)
|
261 |
-
df["answer_len"] = df["answer"].apply(
|
262 |
-
lambda x: len(x) if isinstance(x, str) else 0
|
263 |
-
)
|
264 |
-
|
265 |
-
df[["adjusted_precision", "adjusted_recall"]] = df.apply(
|
266 |
-
calculate_adjusted_performance, axis=1
|
267 |
-
)
|
268 |
-
|
269 |
-
df.to_csv(result_file, index=False)
|
270 |
-
print(f"performance scores saved to result file: {result_file}")
|
271 |
-
|
272 |
-
# print(f"df len: {len(df)}")
|
273 |
-
|
274 |
-
return df
|
275 |
-
|
276 |
-
|
277 |
-
def adjust_perf_scores_with_repetition_penalty(result, precision, recall):
|
278 |
-
newline_score = [
|
279 |
-
df["newline_score"].mean() for df in result["df_list_repetition_penalty"]
|
280 |
-
]
|
281 |
-
|
282 |
-
repetition_score = [
|
283 |
-
df["repetition_score"].mean() for df in result["df_list_repetition_penalty"]
|
284 |
-
]
|
285 |
-
|
286 |
-
answer_len = [
|
287 |
-
df["answer_len"].mean() for df in result["df_list_repetition_penalty"]
|
288 |
-
]
|
289 |
-
|
290 |
-
precision = [
|
291 |
-
calc_adjusted_performance(f, n + r, l)
|
292 |
-
for f, n, r, l in zip(precision, newline_score, repetition_score, answer_len)
|
293 |
-
]
|
294 |
-
recall = [
|
295 |
-
calc_adjusted_performance(f, n + r, l)
|
296 |
-
for f, n, r, l in zip(recall, newline_score, repetition_score, answer_len)
|
297 |
-
]
|
298 |
-
|
299 |
-
return precision, recall
|
300 |
-
|
301 |
-
|
302 |
-
def plot_performance_scores(
|
303 |
-
result,
|
304 |
-
models=None,
|
305 |
-
title="Performance",
|
306 |
-
):
|
307 |
-
if models is None:
|
308 |
-
models = result.keys()
|
309 |
-
for model in models:
|
310 |
-
print(f"model: {model}")
|
311 |
-
df = result[model]["df_overall"]
|
312 |
-
|
313 |
-
# Calculate the statistics
|
314 |
-
precision = [
|
315 |
-
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
316 |
-
]
|
317 |
-
recall = [
|
318 |
-
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
319 |
-
]
|
320 |
-
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
321 |
-
best_f1 = max(f1)
|
322 |
-
best_f1_index = f1.index(best_f1)
|
323 |
-
|
324 |
-
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
325 |
-
result[model], precision, recall
|
326 |
-
)
|
327 |
-
afrp = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
328 |
-
|
329 |
-
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
330 |
-
best_afrp = max(afrp)
|
331 |
-
best_afrp_index = afrp.index(best_afrp)
|
332 |
-
|
333 |
-
adjusted_precision = [
|
334 |
-
df["adjusted_precision"].mean()
|
335 |
-
for df in result[model]["df_list_repetition_penalty"]
|
336 |
-
]
|
337 |
-
adjusted_recall = [
|
338 |
-
df["adjusted_recall"].mean()
|
339 |
-
for df in result[model]["df_list_repetition_penalty"]
|
340 |
-
]
|
341 |
-
afrp2 = [
|
342 |
-
2 * (p * r) / (p + r) for p, r in zip(adjusted_precision, adjusted_recall)
|
343 |
-
]
|
344 |
-
best_afrp2 = max(afrp2)
|
345 |
-
best_afrp2_index = afrp2.index(best_afrp2)
|
346 |
-
|
347 |
-
repetition_penalties = list(df["repetition_penalty"])
|
348 |
-
|
349 |
-
# line plot for precision, recall, f1
|
350 |
-
plt.figure(figsize=(10, 6))
|
351 |
-
|
352 |
-
plt.axvspan(
|
353 |
-
repetition_penalties[best_f1_index] - 0.01,
|
354 |
-
repetition_penalties[best_f1_index] + 0.01,
|
355 |
-
alpha=0.5,
|
356 |
-
edgecolor="none",
|
357 |
-
facecolor="blue",
|
358 |
-
)
|
359 |
-
|
360 |
-
# plt.axvspan(
|
361 |
-
# repetition_penalties[best_afrp2_index] - 0.01,
|
362 |
-
# repetition_penalties[best_afrp2_index] + 0.01,
|
363 |
-
# alpha=0.5,
|
364 |
-
# edgecolor="none",
|
365 |
-
# facecolor="green",
|
366 |
-
# )
|
367 |
-
|
368 |
-
plt.axvspan(
|
369 |
-
repetition_penalties[best_afrp_index] - 0.01,
|
370 |
-
repetition_penalties[best_afrp_index] + 0.01,
|
371 |
-
alpha=0.5,
|
372 |
-
edgecolor="none",
|
373 |
-
facecolor="orange",
|
374 |
-
)
|
375 |
-
|
376 |
-
plt.plot(repetition_penalties, f1, label="F1", marker="D", color="blue")
|
377 |
-
# plt.plot(
|
378 |
-
# repetition_penalties,
|
379 |
-
# afrp2,
|
380 |
-
# label="Per-question RAP - F1",
|
381 |
-
# marker="s",
|
382 |
-
# color="green",
|
383 |
-
# )
|
384 |
-
plt.plot(
|
385 |
-
repetition_penalties,
|
386 |
-
afrp,
|
387 |
-
label="RAP - F1",
|
388 |
-
marker="o",
|
389 |
-
color="orange",
|
390 |
-
)
|
391 |
-
plt.xlabel("Repetition Penalties")
|
392 |
-
plt.ylabel("Score")
|
393 |
-
# plt.xlim(0.99, 1.31)
|
394 |
-
# y in percentage
|
395 |
-
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
396 |
-
plt.title(f"{model} {title}")
|
397 |
-
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
398 |
-
|
399 |
-
plt.show()
|
400 |
-
|
401 |
-
|
402 |
-
def plot_best_afrp(
|
403 |
-
result,
|
404 |
-
models=None,
|
405 |
-
title="Models with Best RAP - F1",
|
406 |
-
ref_result=None,
|
407 |
-
):
|
408 |
-
# Initialize lists to store the statistics
|
409 |
-
model_names = []
|
410 |
-
best_f1 = []
|
411 |
-
best_afrp = []
|
412 |
-
best_repetition_penalty = []
|
413 |
-
best_mtr = []
|
414 |
-
|
415 |
-
if models is None:
|
416 |
-
models = result.keys()
|
417 |
-
for model in models:
|
418 |
-
print(f"model: {model}")
|
419 |
-
df = result[model]["df_overall"]
|
420 |
-
|
421 |
-
# Calculate the statistics
|
422 |
-
precision = [
|
423 |
-
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
424 |
-
]
|
425 |
-
recall = [
|
426 |
-
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
427 |
-
]
|
428 |
-
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
429 |
-
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
430 |
-
|
431 |
-
newline_score = [
|
432 |
-
df["newline_score"].mean()
|
433 |
-
for df in result[model]["df_list_repetition_penalty"]
|
434 |
-
]
|
435 |
-
# print(f"newline_score: {newline_score}")
|
436 |
-
|
437 |
-
repetition_score = [
|
438 |
-
df["repetition_score"].mean()
|
439 |
-
for df in result[model]["df_list_repetition_penalty"]
|
440 |
-
]
|
441 |
-
# print(f"repetition_score: {repetition_score}")
|
442 |
-
|
443 |
-
answer_len = [
|
444 |
-
df["answer_len"].mean()
|
445 |
-
for df in result[model]["df_list_repetition_penalty"]
|
446 |
-
]
|
447 |
-
|
448 |
-
afrp = [
|
449 |
-
calc_adjusted_performance(f, n + r, l)
|
450 |
-
for f, n, r, l in zip(f1, newline_score, repetition_score, answer_len)
|
451 |
-
]
|
452 |
-
|
453 |
-
best_afrp.append(max(afrp))
|
454 |
-
best_afrp_index = afrp.index(best_afrp[-1])
|
455 |
-
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
456 |
-
|
457 |
-
best_f1.append(f1[best_afrp_index])
|
458 |
-
best_mtr.append(
|
459 |
-
newline_score[best_afrp_index] + repetition_score[best_afrp_index]
|
460 |
-
)
|
461 |
-
|
462 |
-
# print(
|
463 |
-
# f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
464 |
-
# )
|
465 |
-
|
466 |
-
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
467 |
-
|
468 |
-
model_names.append(
|
469 |
-
f"{model} (RP={best_repetition_penalty[-1]})"
|
470 |
-
) # Add the model name to the list
|
471 |
-
|
472 |
-
if ref_result is not None:
|
473 |
-
print("ref_result:", ref_result)
|
474 |
-
for model in ref_result.keys():
|
475 |
-
model_names.append(model)
|
476 |
-
df = pd.read_csv(ref_result[model])
|
477 |
-
# df = df[df["id"].isin(wikidata_df["id"])]
|
478 |
-
|
479 |
-
p = df["precision"].mean()
|
480 |
-
r = df["recall"].mean()
|
481 |
-
|
482 |
-
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
483 |
-
best_f1.append(f1)
|
484 |
-
best_afrp.append(f1)
|
485 |
-
best_mtr.append(0)
|
486 |
-
|
487 |
-
print("model_names:", model_names)
|
488 |
-
# print("best_f1:", best_f1)
|
489 |
-
# print("best_afrp:", best_afrp)
|
490 |
-
|
491 |
-
# Create a DataFrame with the statistics
|
492 |
-
data = pd.DataFrame(
|
493 |
-
{
|
494 |
-
"Model": model_names,
|
495 |
-
"RAP - F1": best_afrp,
|
496 |
-
"F1": best_f1,
|
497 |
-
}
|
498 |
-
)
|
499 |
-
|
500 |
-
# Melt the DataFrame to a long format
|
501 |
-
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
502 |
-
|
503 |
-
# Pivot the DataFrame to a wide format
|
504 |
-
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
505 |
-
|
506 |
-
# make sure the columns are following the order of the models
|
507 |
-
data_pivoted = data_pivoted[model_names]
|
508 |
-
|
509 |
-
# make sure three groups in the order of precision, recall, f1
|
510 |
-
data_pivoted = data_pivoted.reindex(["RAP - F1", "F1"])
|
511 |
-
|
512 |
-
# Plot the statistics
|
513 |
-
plt.figure(figsize=(15, 6))
|
514 |
-
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
515 |
-
plt.title(title)
|
516 |
-
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
517 |
-
|
518 |
-
# Set the rotation of the x-axis labels to 0 degrees
|
519 |
-
plt.xticks(rotation=0)
|
520 |
-
|
521 |
-
# Format the y-axis to display as percentage
|
522 |
-
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
523 |
-
|
524 |
-
# get the max value of the y-axis
|
525 |
-
a1 = max(best_afrp)
|
526 |
-
a2 = max(best_f1)
|
527 |
-
|
528 |
-
max_value = max([a1, a2]) * 1.12
|
529 |
-
print("max_value:", max_value)
|
530 |
-
|
531 |
-
# Set the y-axis limit up to 70%
|
532 |
-
ax.set_ylim(0, max_value)
|
533 |
-
|
534 |
-
# Add the values above each bar
|
535 |
-
for p in ax.patches:
|
536 |
-
ax.annotate(
|
537 |
-
f"{p.get_height() * 100:.1f}",
|
538 |
-
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
539 |
-
ha="center",
|
540 |
-
va="bottom",
|
541 |
-
xytext=(0, 10),
|
542 |
-
textcoords="offset points",
|
543 |
-
rotation=90,
|
544 |
-
)
|
545 |
-
|
546 |
-
plt.show()
|
547 |
-
return data_pivoted, best_mtr
|
548 |
-
|
549 |
-
|
550 |
-
def plot_best_performance(
|
551 |
-
result,
|
552 |
-
models=None,
|
553 |
-
title="Models with Best F1 Score",
|
554 |
-
adjusted_f1=False,
|
555 |
-
ref_result=None,
|
556 |
-
):
|
557 |
-
# Initialize lists to store the statistics
|
558 |
-
model_names = []
|
559 |
-
best_precision = []
|
560 |
-
best_recall = []
|
561 |
-
best_f1 = []
|
562 |
-
best_repetition_penalty = []
|
563 |
-
best_mtr = []
|
564 |
-
|
565 |
-
if models is None:
|
566 |
-
models = result.keys()
|
567 |
-
for model in models:
|
568 |
-
print(f"model: {model}")
|
569 |
-
df = result[model]["df_overall"]
|
570 |
-
|
571 |
-
# Calculate the statistics
|
572 |
-
precision = [
|
573 |
-
df["precision"].mean() for df in result[model]["df_list_repetition_penalty"]
|
574 |
-
]
|
575 |
-
recall = [
|
576 |
-
df["recall"].mean() for df in result[model]["df_list_repetition_penalty"]
|
577 |
-
]
|
578 |
-
newline_score = [
|
579 |
-
df["newline_score"].mean()
|
580 |
-
for df in result[model]["df_list_repetition_penalty"]
|
581 |
-
]
|
582 |
-
|
583 |
-
repetition_score = [
|
584 |
-
df["repetition_score"].mean()
|
585 |
-
for df in result[model]["df_list_repetition_penalty"]
|
586 |
-
]
|
587 |
-
|
588 |
-
if adjusted_f1:
|
589 |
-
precision, recall = adjust_perf_scores_with_repetition_penalty(
|
590 |
-
result[model], precision, recall
|
591 |
-
)
|
592 |
-
|
593 |
-
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
594 |
-
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
595 |
-
|
596 |
-
best_f1.append(max(f1))
|
597 |
-
best_f1_index = f1.index(best_f1[-1])
|
598 |
-
best_repetition_penalty.append(df["repetition_penalty"][best_f1_index])
|
599 |
-
|
600 |
-
best_precision.append(precision[best_f1_index])
|
601 |
-
best_recall.append(recall[best_f1_index])
|
602 |
-
best_mtr.append(newline_score[best_f1_index] + repetition_score[best_f1_index])
|
603 |
-
|
604 |
-
print(
|
605 |
-
f"best repetition penalty: {best_repetition_penalty[-1]}, best f1: {best_f1[-1]}, precision: {best_precision[-1]}, recall: {best_recall[-1]}"
|
606 |
-
)
|
607 |
-
|
608 |
-
df = result[model]["df_list_repetition_penalty"][best_f1_index]
|
609 |
-
|
610 |
-
model_names.append(
|
611 |
-
f"{model} (RP={best_repetition_penalty[-1]})"
|
612 |
-
) # Add the model name to the list
|
613 |
-
|
614 |
-
# print sum for columns: newline_score, repetition_score
|
615 |
-
print(
|
616 |
-
f"newline_score: {df['newline_score'].sum()}, repetition_score: {df['repetition_score'].sum()}"
|
617 |
-
)
|
618 |
-
|
619 |
-
if ref_result is not None:
|
620 |
-
print("ref_result:", ref_result)
|
621 |
-
for model in ref_result.keys():
|
622 |
-
model_names.append(model)
|
623 |
-
df = pd.read_csv(ref_result[model])
|
624 |
-
# df = df[df["id"].isin(wikidata_df["id"])]
|
625 |
-
|
626 |
-
best_precision.append(df["precision"].mean())
|
627 |
-
best_recall.append(df["recall"].mean())
|
628 |
-
f1 = (
|
629 |
-
2
|
630 |
-
* (best_precision[-1] * best_recall[-1])
|
631 |
-
/ (best_precision[-1] + best_recall[-1])
|
632 |
-
)
|
633 |
-
# best_f1.append(df["f1"].mean())
|
634 |
-
best_f1.append(f1)
|
635 |
-
best_mtr.append(0)
|
636 |
-
|
637 |
-
# Create a DataFrame with the statistics
|
638 |
-
data = (
|
639 |
-
pd.DataFrame(
|
640 |
-
{
|
641 |
-
"Model": model_names,
|
642 |
-
"Adjusted Precision with RP": best_precision,
|
643 |
-
"Adjusted Recall with RP": best_recall,
|
644 |
-
"Adjusted F1 with RP": best_f1,
|
645 |
-
}
|
646 |
-
)
|
647 |
-
if adjusted_f1
|
648 |
-
else pd.DataFrame(
|
649 |
-
{
|
650 |
-
"Model": model_names,
|
651 |
-
"Precision": best_precision,
|
652 |
-
"Recall": best_recall,
|
653 |
-
"F1": best_f1,
|
654 |
-
}
|
655 |
-
)
|
656 |
-
)
|
657 |
-
columns = list(data.columns)
|
658 |
-
|
659 |
-
# Melt the DataFrame to a long format
|
660 |
-
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
661 |
-
|
662 |
-
# Pivot the DataFrame to a wide format
|
663 |
-
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
664 |
-
|
665 |
-
# make sure the columns are following the order of the models
|
666 |
-
data_pivoted = data_pivoted[model_names]
|
667 |
-
|
668 |
-
# make sure three groups in the order of precision, recall, f1
|
669 |
-
data_pivoted = data_pivoted.reindex(columns[1:])
|
670 |
-
|
671 |
-
# Plot the statistics
|
672 |
-
plt.figure(figsize=(10, 6))
|
673 |
-
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
674 |
-
plt.title(title)
|
675 |
-
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
676 |
-
|
677 |
-
# Set the rotation of the x-axis labels to 0 degrees
|
678 |
-
plt.xticks(rotation=0)
|
679 |
-
|
680 |
-
# Format the y-axis to display as percentage
|
681 |
-
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
682 |
-
|
683 |
-
# get the max value of the y-axis
|
684 |
-
a1 = max(best_precision)
|
685 |
-
a2 = max(best_recall)
|
686 |
-
a3 = max(best_f1)
|
687 |
-
|
688 |
-
max_value = max([a1, a2, a3]) * 1.12
|
689 |
-
print("max_value:", max_value)
|
690 |
-
|
691 |
-
# Set the y-axis limit up to 70%
|
692 |
-
ax.set_ylim(0, max_value)
|
693 |
-
|
694 |
-
# Add the values above each bar
|
695 |
-
for p in ax.patches:
|
696 |
-
ax.annotate(
|
697 |
-
f"{p.get_height() * 100:.1f}",
|
698 |
-
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
699 |
-
ha="center",
|
700 |
-
va="bottom",
|
701 |
-
xytext=(0, 10),
|
702 |
-
textcoords="offset points",
|
703 |
-
rotation=90,
|
704 |
-
)
|
705 |
-
|
706 |
-
plt.show()
|
707 |
-
return data_pivoted, best_mtr
|
708 |
-
|
709 |
-
|
710 |
-
def plot_best_performance_ms_macro(
|
711 |
-
result,
|
712 |
-
models=None,
|
713 |
-
title="Models with Best RAP - Performance",
|
714 |
-
ref_result=None,
|
715 |
-
skip_generic_prompt=False,
|
716 |
-
include_adjusted_performance=True,
|
717 |
-
):
|
718 |
-
# Initialize lists to store the statistics
|
719 |
-
model_names = []
|
720 |
-
best_f1 = []
|
721 |
-
best_afrp = []
|
722 |
-
best_repetition_penalty = []
|
723 |
-
best_bleu1 = []
|
724 |
-
best_rougeL = []
|
725 |
-
best_mtr = []
|
726 |
-
|
727 |
-
if models is None:
|
728 |
-
models = result.keys()
|
729 |
-
for model in models:
|
730 |
-
if skip_generic_prompt and "generic prompt" in model:
|
731 |
-
continue
|
732 |
-
print(f"model: {model}")
|
733 |
-
df = result[model]["df_overall"]
|
734 |
-
|
735 |
-
# Calculate the statistics
|
736 |
-
bleu1 = [x for x in df["bleu1"]]
|
737 |
-
rougeL = [x for x in df["rougeL"]]
|
738 |
-
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
739 |
-
|
740 |
-
newline_score = [
|
741 |
-
df["newline_score"].mean()
|
742 |
-
for df in result[model]["df_list_repetition_penalty"]
|
743 |
-
]
|
744 |
-
# print(f"newline_score: {newline_score}")
|
745 |
-
|
746 |
-
repetition_score = [
|
747 |
-
df["repetition_score"].mean()
|
748 |
-
for df in result[model]["df_list_repetition_penalty"]
|
749 |
-
]
|
750 |
-
# print(f"repetition_score: {repetition_score}")
|
751 |
-
|
752 |
-
answer_len = [
|
753 |
-
df["answer_len"].mean()
|
754 |
-
for df in result[model]["df_list_repetition_penalty"]
|
755 |
-
]
|
756 |
-
|
757 |
-
afrp = [
|
758 |
-
calc_adjusted_performance(f, n + r, l)
|
759 |
-
for f, n, r, l in zip(f1, newline_score, repetition_score, answer_len)
|
760 |
-
]
|
761 |
-
|
762 |
-
best_afrp.append(max(afrp if include_adjusted_performance else f1))
|
763 |
-
best_afrp_index = (
|
764 |
-
afrp.index(best_afrp[-1])
|
765 |
-
if include_adjusted_performance
|
766 |
-
else f1.index(best_afrp[-1])
|
767 |
-
)
|
768 |
-
best_repetition_penalty.append(df["repetition_penalty"][best_afrp_index])
|
769 |
-
|
770 |
-
best_f1.append(f1[best_afrp_index])
|
771 |
-
best_bleu1.append(bleu1[best_afrp_index])
|
772 |
-
best_rougeL.append(rougeL[best_afrp_index])
|
773 |
-
best_mtr.append(
|
774 |
-
newline_score[best_afrp_index] + repetition_score[best_afrp_index]
|
775 |
-
)
|
776 |
-
|
777 |
-
# print(
|
778 |
-
# f"best repetition penalty: {best_repetition_penalty[-1]}, best afrp: {best_afrp[-1]}, f1: {best_f1[-1]}"
|
779 |
-
# )
|
780 |
-
|
781 |
-
df = result[model]["df_list_repetition_penalty"][best_afrp_index]
|
782 |
-
|
783 |
-
model_names.append(
|
784 |
-
f"{model} (RP={best_repetition_penalty[-1]})"
|
785 |
-
) # Add the model name to the list
|
786 |
-
|
787 |
-
if ref_result is not None:
|
788 |
-
print("ref_result:", ref_result)
|
789 |
-
for model in ref_result.keys():
|
790 |
-
model_names.append(model)
|
791 |
-
df = pd.read_csv(ref_result[model], comment="#", on_bad_lines="warn")
|
792 |
-
# df = df[df["id"].isin(wikidata_df["id"])]
|
793 |
-
|
794 |
-
p = df["bleu1"][0]
|
795 |
-
best_bleu1.append(p)
|
796 |
-
|
797 |
-
r = df["rougeL"][0]
|
798 |
-
best_rougeL.append(r)
|
799 |
-
|
800 |
-
f1 = 2 * p * r / (p + r) if p + r > 0 else 0
|
801 |
-
best_f1.append(f1)
|
802 |
-
best_afrp.append(f1)
|
803 |
-
best_mtr.append(0)
|
804 |
-
|
805 |
-
# print("model_names:", model_names)
|
806 |
-
# print("best_f1:", best_f1)
|
807 |
-
# print("best_afrp:", best_afrp)
|
808 |
-
|
809 |
-
# Create a DataFrame with the statistics
|
810 |
-
data = (
|
811 |
-
pd.DataFrame(
|
812 |
-
{
|
813 |
-
"Model": model_names,
|
814 |
-
"RAP - Perf Score": best_afrp,
|
815 |
-
"Overall Perf Score": best_f1,
|
816 |
-
}
|
817 |
-
)
|
818 |
-
if include_adjusted_performance
|
819 |
-
else pd.DataFrame(
|
820 |
-
{
|
821 |
-
"Model": model_names,
|
822 |
-
"Bleu-1": best_bleu1,
|
823 |
-
"Rouge-L": best_rougeL,
|
824 |
-
"Overall Perf Score": best_f1,
|
825 |
-
}
|
826 |
-
)
|
827 |
-
)
|
828 |
-
|
829 |
-
# Melt the DataFrame to a long format
|
830 |
-
data_melted = data.melt(id_vars="Model", var_name="Metric", value_name="Score")
|
831 |
-
|
832 |
-
# Pivot the DataFrame to a wide format
|
833 |
-
data_pivoted = data_melted.pivot(index="Metric", columns="Model", values="Score")
|
834 |
-
|
835 |
-
# make sure the columns are following the order of the models
|
836 |
-
data_pivoted = data_pivoted[model_names]
|
837 |
-
|
838 |
-
columns = list(data.columns)
|
839 |
-
data_pivoted = data_pivoted.reindex(columns[1:])
|
840 |
-
|
841 |
-
# Plot the statistics
|
842 |
-
plt.figure(figsize=(10, 6))
|
843 |
-
ax = data_pivoted.plot(kind="bar", ax=plt.gca(), width=0.9)
|
844 |
-
plt.title(title)
|
845 |
-
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
846 |
-
|
847 |
-
# Set the rotation of the x-axis labels to 0 degrees
|
848 |
-
plt.xticks(rotation=0)
|
849 |
-
|
850 |
-
# Format the y-axis to display as percentage
|
851 |
-
ax.yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
852 |
-
|
853 |
-
# get the max value of the y-axis
|
854 |
-
a1 = max(best_afrp)
|
855 |
-
a2 = max(best_f1)
|
856 |
-
a3 = max(best_bleu1)
|
857 |
-
a4 = max(best_rougeL)
|
858 |
-
|
859 |
-
max_value = (
|
860 |
-
max([a1, a2] if include_adjusted_performance else [a1, a2, a3, a4]) * 1.12
|
861 |
-
)
|
862 |
-
print("max_value:", max_value)
|
863 |
-
|
864 |
-
# Set the y-axis limit up to 70%
|
865 |
-
ax.set_ylim(0, max_value)
|
866 |
-
|
867 |
-
# Add the values above each bar
|
868 |
-
for p in ax.patches:
|
869 |
-
ax.annotate(
|
870 |
-
f"{p.get_height() * 100:.1f}",
|
871 |
-
(p.get_x() + p.get_width() / 2.0, p.get_height()),
|
872 |
-
ha="center",
|
873 |
-
va="bottom",
|
874 |
-
xytext=(0, 10),
|
875 |
-
textcoords="offset points",
|
876 |
-
rotation=90,
|
877 |
-
)
|
878 |
-
|
879 |
-
plt.show()
|
880 |
-
return data_pivoted, best_mtr
|
881 |
-
|
882 |
-
|
883 |
-
all_open_source_models = [
|
884 |
-
"gemma-1.1-2b-it",
|
885 |
-
"Phi-3-mini-128k-instruct",
|
886 |
-
"gemma-1.1-7b-it",
|
887 |
-
"Llama-2-7b-chat-hf",
|
888 |
-
"Mistral-7B-Instruct-v0.2",
|
889 |
-
"Meta-Llama-3-8B-Instruct",
|
890 |
-
"Llama-2-13b-chat-hf",
|
891 |
-
"Llama-2-70b-chat-hf",
|
892 |
-
"Meta-Llama-3-70B-Instruct",
|
893 |
-
]
|
894 |
-
|
895 |
-
|
896 |
-
def load_for_repetition_penalty_ms_macro(
|
897 |
-
csv_result_file, repetition_penalty, force_recalculate=False
|
898 |
-
):
|
899 |
-
result_file = replace_last(
|
900 |
-
csv_result_file, ".csv", f"_rpp_{repetition_penalty:.2f}.csv"
|
901 |
-
)
|
902 |
-
df = load_with_newline_and_repetition_scores(
|
903 |
-
result_file, force_recalculate=force_recalculate
|
904 |
-
)
|
905 |
-
|
906 |
-
return df
|
907 |
-
|
908 |
-
|
909 |
-
# MS MACRO
|
910 |
-
def plot_performance_scores_ms_macro(
|
911 |
-
result,
|
912 |
-
models=None,
|
913 |
-
title="Performance",
|
914 |
-
):
|
915 |
-
if models is None:
|
916 |
-
models = result.keys()
|
917 |
-
for model in models:
|
918 |
-
print(f"model: {model}")
|
919 |
-
df = result[model]["df_overall"]
|
920 |
-
# print(result[model]["df_list_repetition_penalty"][0].describe())
|
921 |
-
|
922 |
-
# Calculate the statistics
|
923 |
-
bleu1 = list(df["bleu1"])
|
924 |
-
rougeL = list(df["rougeL"])
|
925 |
-
f1 = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
926 |
-
best_f1 = max(f1)
|
927 |
-
best_f1_index = f1.index(best_f1)
|
928 |
-
|
929 |
-
bleu1, rougeL = adjust_perf_scores_with_repetition_penalty(
|
930 |
-
result[model], bleu1, rougeL
|
931 |
-
)
|
932 |
-
afrp = [2 * (p * r) / (p + r) for p, r in zip(bleu1, rougeL)]
|
933 |
-
|
934 |
-
# f1 = [df["f1"].mean() for df in result[model]["df_list_repetition_penalty"]]
|
935 |
-
best_afrp = max(afrp)
|
936 |
-
best_afrp_index = afrp.index(best_afrp)
|
937 |
-
|
938 |
-
repetition_penalties = list(df["repetition_penalty"])
|
939 |
-
|
940 |
-
# line plot for precision, recall, f1
|
941 |
-
plt.figure(figsize=(10, 6))
|
942 |
-
|
943 |
-
plt.axvspan(
|
944 |
-
repetition_penalties[best_f1_index] - 0.01,
|
945 |
-
repetition_penalties[best_f1_index] + 0.01,
|
946 |
-
alpha=0.5,
|
947 |
-
edgecolor="none",
|
948 |
-
facecolor="blue",
|
949 |
-
)
|
950 |
-
|
951 |
-
plt.axvspan(
|
952 |
-
repetition_penalties[best_afrp_index] - 0.01,
|
953 |
-
repetition_penalties[best_afrp_index] + 0.01,
|
954 |
-
alpha=0.5,
|
955 |
-
edgecolor="none",
|
956 |
-
facecolor="orange",
|
957 |
-
)
|
958 |
-
|
959 |
-
plt.plot(
|
960 |
-
repetition_penalties,
|
961 |
-
f1,
|
962 |
-
label="Overall Perf Score",
|
963 |
-
marker="D",
|
964 |
-
color="blue",
|
965 |
-
)
|
966 |
-
plt.plot(
|
967 |
-
repetition_penalties,
|
968 |
-
afrp,
|
969 |
-
label="RAP - Perf Score",
|
970 |
-
marker="o",
|
971 |
-
color="orange",
|
972 |
-
)
|
973 |
-
|
974 |
-
plt.xlabel("Repetition Penalties")
|
975 |
-
plt.ylabel("Score")
|
976 |
-
# plt.xlim(0.99, 1.31)
|
977 |
-
# y in percentage
|
978 |
-
plt.gca().yaxis.set_major_formatter(mtick.PercentFormatter(1.0))
|
979 |
-
plt.title(f"{model} {title}")
|
980 |
-
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
981 |
-
|
982 |
-
plt.show()
|
983 |
-
|
984 |
-
|
985 |
-
def plot_repetition_factors(result, groups):
|
986 |
-
for group in groups:
|
987 |
-
# Plot the statistics
|
988 |
-
plt.figure(figsize=(10, 6))
|
989 |
-
|
990 |
-
max_value = 0
|
991 |
-
for model in result.keys():
|
992 |
-
if not group in model.lower():
|
993 |
-
continue
|
994 |
-
print(f"model: {model}")
|
995 |
-
df = result[model]["df_overall"]
|
996 |
-
repetition_panelties = [
|
997 |
-
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
998 |
-
]
|
999 |
-
|
1000 |
-
mean_score = [
|
1001 |
-
df["total_repetitions"].mean()
|
1002 |
-
for df in result[model]["df_list_repetition_penalty"]
|
1003 |
-
]
|
1004 |
-
|
1005 |
-
sns.lineplot(x=repetition_panelties, y=mean_score, label=model)
|
1006 |
-
|
1007 |
-
new_max = max(mean_score)
|
1008 |
-
if new_max > max_value:
|
1009 |
-
max_value = new_max
|
1010 |
-
|
1011 |
-
max_value = max_value * 1.05
|
1012 |
-
# if max_value < 1.5:
|
1013 |
-
# max_value = 1.5
|
1014 |
-
# set ylimit
|
1015 |
-
plt.ylim(0, max_value)
|
1016 |
-
|
1017 |
-
# show grid
|
1018 |
-
plt.grid(True)
|
1019 |
-
plt.xlabel("Repetition Penalties")
|
1020 |
-
plt.ylabel("Mean Total Repetitions")
|
1021 |
-
plt.title("Mean Total Repetitions vs Repetition Penalties")
|
1022 |
-
plt.legend()
|
1023 |
-
|
1024 |
-
plt.show()
|
1025 |
-
|
1026 |
-
|
1027 |
-
def plot_repetition_factors_by_group(result, group_filter=None):
|
1028 |
-
markers = ["D", "o", "s", "x"]
|
1029 |
-
colors = ["blue", "orange", "green", "red"]
|
1030 |
-
|
1031 |
-
# Plot the statistics
|
1032 |
-
plt.figure(figsize=(10, 6))
|
1033 |
-
index = 0
|
1034 |
-
max_value = 0
|
1035 |
-
|
1036 |
-
for model in result.keys():
|
1037 |
-
if group_filter is not None and group_filter not in model:
|
1038 |
-
continue
|
1039 |
-
|
1040 |
-
print(f"model: {model}")
|
1041 |
-
|
1042 |
-
df = result[model]["df_overall"]
|
1043 |
-
repetition_panelties = [
|
1044 |
-
repetition_penalty for repetition_penalty in df["repetition_penalty"]
|
1045 |
-
]
|
1046 |
-
|
1047 |
-
# Calculate the statistics
|
1048 |
-
mean_score = [
|
1049 |
-
df["total_repetitions"].mean()
|
1050 |
-
for df in result[model]["df_list_repetition_penalty"]
|
1051 |
-
]
|
1052 |
-
if len(mean_score) != len(repetition_panelties):
|
1053 |
-
print(
|
1054 |
-
f"model: {model} has different length of repetition penalties and mean score"
|
1055 |
-
)
|
1056 |
-
print("repetition_panelties:", len(repetition_panelties))
|
1057 |
-
print("mean_score:", len(mean_score))
|
1058 |
-
continue
|
1059 |
-
|
1060 |
-
new_max = max(mean_score)
|
1061 |
-
if new_max > max_value:
|
1062 |
-
max_value = new_max
|
1063 |
-
|
1064 |
-
sns.lineplot(
|
1065 |
-
x=repetition_panelties,
|
1066 |
-
y=mean_score,
|
1067 |
-
label=model,
|
1068 |
-
marker=markers[index],
|
1069 |
-
color=colors[index],
|
1070 |
-
)
|
1071 |
-
|
1072 |
-
index += 1
|
1073 |
-
|
1074 |
-
max_value = max_value * 1.05
|
1075 |
-
# if max_value < 1.5:
|
1076 |
-
# max_value = 1.5
|
1077 |
-
# set ylimit
|
1078 |
-
plt.ylim(0, max_value)
|
1079 |
-
max_value = 0
|
1080 |
-
|
1081 |
-
plt.xlabel("Repetition Penalties")
|
1082 |
-
plt.ylabel("Mean Total Repetitions")
|
1083 |
-
plt.title("Mean Total Repetitions vs Repetition Penalties")
|
1084 |
-
plt.legend(bbox_to_anchor=(1.0, 0.5), loc="center left")
|
1085 |
-
|
1086 |
-
plt.show()
|
1087 |
-
|
1088 |
-
|
1089 |
-
ms_marco_csv_result_files = [
|
1090 |
-
"data/results_v2/gemma-1.1-2b-it(RAG - Generic Prompt)_mm.csv",
|
1091 |
-
"data/results_v2/gemma-1.1-2b-it(RAG - Chat Template)_mm.csv",
|
1092 |
-
"data/results_v2/gemma-1.1-2b-it(Non-RAG)_mm.csv",
|
1093 |
-
"data/results_v2/Phi-3-mini-128k-instruct(RAG - Generic Prompt)_mm.csv",
|
1094 |
-
"data/results_v2/Phi-3-mini-128k-instruct(RAG - Chat Template)_mm.csv",
|
1095 |
-
"data/results_v2/Phi-3-mini-128k-instruct(Non-RAG)_mm.csv",
|
1096 |
-
"data/results_v2/gemma-1.1-7b-it(RAG - Generic Prompt)_mm.csv",
|
1097 |
-
"data/results_v2/gemma-1.1-7b-it(RAG - Chat Template)_mm.csv",
|
1098 |
-
"data/results_v2/gemma-1.1-7b-it(Non-RAG)_mm.csv",
|
1099 |
-
"data/results_v2/Llama-2-7b-chat-hf(RAG - Generic Prompt)_mm.csv",
|
1100 |
-
"data/results_v2/Llama-2-7b-chat-hf(RAG - Chat Template)_mm.csv",
|
1101 |
-
"data/results_v2/Llama-2-7b-chat-hf(Non-RAG)_mm.csv",
|
1102 |
-
"data/results_v2/Mistral-7B-Instruct-v0.2(RAG - Generic Prompt)_mm.csv",
|
1103 |
-
"data/results_v2/Mistral-7B-Instruct-v0.2(RAG - Chat Template)_mm.csv",
|
1104 |
-
"data/results_v2/Mistral-7B-Instruct-v0.2(Non-RAG)_mm.csv",
|
1105 |
-
"data/results_v2/Meta-Llama-3-8B-Instruct(RAG - Generic Prompt)_mm.csv",
|
1106 |
-
"data/results_v2/Meta-Llama-3-8B-Instruct(RAG - Chat Template)_mm.csv",
|
1107 |
-
"data/results_v2/Meta-Llama-3-8B-Instruct(Non-RAG)_mm.csv",
|
1108 |
-
"data/results_v2/Llama-2-13b-chat-hf(RAG - Generic Prompt)_mm.csv",
|
1109 |
-
"data/results_v2/Llama-2-13b-chat-hf(RAG - Chat Template)_mm.csv",
|
1110 |
-
"data/results_v2/Llama-2-13b-chat-hf(Non-RAG)_mm.csv",
|
1111 |
-
"data/results_v2/Llama-2-70b-chat-hf(RAG - Generic Prompt)_mm.csv",
|
1112 |
-
"data/results_v2/Llama-2-70b-chat-hf(RAG - Chat Template)_mm.csv",
|
1113 |
-
"data/results_v2/Llama-2-70b-chat-hf(Non-RAG)_mm.csv",
|
1114 |
-
"data/results_v2/Meta-Llama-3-70B-Instruct(RAG - Generic Prompt)_mm.csv",
|
1115 |
-
"data/results_v2/Meta-Llama-3-70B-Instruct(RAG - Chat Template)_mm.csv",
|
1116 |
-
"data/results_v2/Meta-Llama-3-70B-Instruct(Non-RAG)_mm.csv",
|
1117 |
-
]
|
1118 |
-
|
1119 |
-
webqsp_csv_result_files = [
|
1120 |
-
"data/results_v2/gemma-1.1-2b-it(RAG - Generic Prompt)_wd.csv",
|
1121 |
-
"data/results_v2/gemma-1.1-2b-it(RAG - Chat Template)_wd.csv",
|
1122 |
-
"data/results_v2/gemma-1.1-2b-it(Non-RAG)_wd.csv",
|
1123 |
-
"data/results_v2/Phi-3-mini-128k-instruct(RAG - Generic Prompt)_wd.csv",
|
1124 |
-
"data/results_v2/Phi-3-mini-128k-instruct(RAG - Chat Template)_wd.csv",
|
1125 |
-
"data/results_v2/Phi-3-mini-128k-instruct(Non-RAG)_wd.csv",
|
1126 |
-
"data/results_v2/gemma-1.1-7b-it(RAG - Generic Prompt)_wd.csv",
|
1127 |
-
"data/results_v2/gemma-1.1-7b-it(RAG - Chat Template)_wd.csv",
|
1128 |
-
"data/results_v2/gemma-1.1-7b-it(Non-RAG)_wd.csv",
|
1129 |
-
"data/results_v2/Llama-2-7b-chat-hf(RAG - Generic Prompt)_wd.csv",
|
1130 |
-
"data/results_v2/Llama-2-7b-chat-hf(RAG - Chat Template)_wd.csv",
|
1131 |
-
"data/results_v2/Llama-2-7b-chat-hf(Non-RAG)_wd.csv",
|
1132 |
-
"data/results_v2/Mistral-7B-Instruct-v0.2(RAG - Generic Prompt)_wd.csv",
|
1133 |
-
"data/results_v2/Mistral-7B-Instruct-v0.2(RAG - Chat Template)_wd.csv",
|
1134 |
-
"data/results_v2/Mistral-7B-Instruct-v0.2(Non-RAG)_wd.csv",
|
1135 |
-
"data/results_v2/Meta-Llama-3-8B-Instruct(RAG - Generic Prompt)_wd.csv",
|
1136 |
-
"data/results_v2/Meta-Llama-3-8B-Instruct(RAG - Chat Template)_wd.csv",
|
1137 |
-
"data/results_v2/Meta-Llama-3-8B-Instruct(Non-RAG)_wd.csv",
|
1138 |
-
"data/results_v2/Llama-2-13b-chat-hf(RAG - Generic Prompt)_wd.csv",
|
1139 |
-
"data/results_v2/Llama-2-13b-chat-hf(RAG - Chat Template)_wd.csv",
|
1140 |
-
"data/results_v2/Llama-2-13b-chat-hf(Non-RAG)_wd.csv",
|
1141 |
-
"data/results_v2/Llama-2-70b-chat-hf(RAG - Generic Prompt)_wd.csv",
|
1142 |
-
"data/results_v2/Llama-2-70b-chat-hf(RAG - Chat Template)_wd.csv",
|
1143 |
-
"data/results_v2/Llama-2-70b-chat-hf(Non-RAG)_wd.csv",
|
1144 |
-
"data/results_v2/Meta-Llama-3-70B-Instruct(RAG - Generic Prompt)_wd.csv",
|
1145 |
-
"data/results_v2/Meta-Llama-3-70B-Instruct(RAG - Chat Template)_wd.csv",
|
1146 |
-
"data/results_v2/Meta-Llama-3-70B-Instruct(Non-RAG)_wd.csv",
|
1147 |
-
]
|
1148 |
-
|
1149 |
-
|
1150 |
-
def calc_rap_scores(
|
1151 |
-
result, precision="precision", recall="recall", penalty_function="cubic"
|
1152 |
-
):
|
1153 |
-
newline_score = [
|
1154 |
-
df["newline_score"].mean() for df in result["df_list_repetition_penalty"]
|
1155 |
-
]
|
1156 |
-
|
1157 |
-
repetition_score = [
|
1158 |
-
df["repetition_score"].mean() for df in result["df_list_repetition_penalty"]
|
1159 |
-
]
|
1160 |
-
|
1161 |
-
if precision in result["df_list_repetition_penalty"][0].columns:
|
1162 |
-
precision = [
|
1163 |
-
df[precision].mean() for df in result["df_list_repetition_penalty"]
|
1164 |
-
]
|
1165 |
-
recall = [df[recall].mean() for df in result["df_list_repetition_penalty"]]
|
1166 |
-
else:
|
1167 |
-
precision = result["df_overall"][precision]
|
1168 |
-
recall = result["df_overall"][recall]
|
1169 |
-
|
1170 |
-
f1 = [2 * (p * r) / (p + r) for p, r in zip(precision, recall)]
|
1171 |
-
|
1172 |
-
nrr = [
|
1173 |
-
1 - (n + r) / s
|
1174 |
-
for f, n, r, s in zip(
|
1175 |
-
f1, newline_score, repetition_score, result["df_overall"]["answer_len"]
|
1176 |
-
)
|
1177 |
-
]
|
1178 |
-
|
1179 |
-
rap = [
|
1180 |
-
calc_adjusted_performance(f, 1 - n, penalty_function=penalty_function)
|
1181 |
-
for f, n in zip(f1, nrr)
|
1182 |
-
]
|
1183 |
-
|
1184 |
-
return newline_score, repetition_score, f1, rap, nrr
|
1185 |
-
|
1186 |
-
|
1187 |
-
def get_model_name(csv_result_file):
|
1188 |
-
parts = re.split(r"[_/]", csv_result_file)
|
1189 |
-
print(f"parts: {parts}")
|
1190 |
-
model_name = parts[3]
|
1191 |
-
return model_name
|
1192 |
-
|
1193 |
-
|
1194 |
-
def load_webqsp_result(
|
1195 |
-
csv_result_files, force_recalculate=False, save=False, penalty_function="cubic"
|
1196 |
-
):
|
1197 |
-
result = {}
|
1198 |
-
for i, csv_result_file in enumerate(csv_result_files):
|
1199 |
-
try:
|
1200 |
-
df = pd.read_csv(csv_result_file)
|
1201 |
-
model_name = get_model_name(csv_result_file)
|
1202 |
-
print(f"\tmodel_name: {model_name}")
|
1203 |
-
|
1204 |
-
dfs = [
|
1205 |
-
calculate_performance_score(
|
1206 |
-
csv_result_file,
|
1207 |
-
repetition_penalty,
|
1208 |
-
force_recalculate=force_recalculate,
|
1209 |
-
)
|
1210 |
-
for repetition_penalty in df["repetition_penalty"]
|
1211 |
-
]
|
1212 |
-
|
1213 |
-
answer_lens = []
|
1214 |
-
for df_rpp in dfs:
|
1215 |
-
answer_lens.append(df_rpp["answer_len"].mean())
|
1216 |
-
df["answer_len"] = answer_lens
|
1217 |
-
|
1218 |
-
result[model_name] = {
|
1219 |
-
"df_overall": df,
|
1220 |
-
"df_list_repetition_penalty": dfs,
|
1221 |
-
"file": csv_result_file,
|
1222 |
-
}
|
1223 |
-
newline_score, repetition_score, perf, rap, nrr = calc_rap_scores(
|
1224 |
-
result[model_name], penalty_function=penalty_function
|
1225 |
-
)
|
1226 |
-
df["newline_score"] = newline_score
|
1227 |
-
df["repetition_score"] = repetition_score
|
1228 |
-
df["total_repetitions"] = df["newline_score"] + df["repetition_score"]
|
1229 |
-
df["perf"] = perf
|
1230 |
-
df["nrr"] = nrr
|
1231 |
-
df["rap"] = rap
|
1232 |
-
df["rr"] = df["nrr"].apply(lambda x: 1 - x)
|
1233 |
-
df["rrp"] = df["rr"].apply(lambda x: x * 100)
|
1234 |
-
if save:
|
1235 |
-
df.to_csv(csv_result_file, index=False)
|
1236 |
-
except Exception as e:
|
1237 |
-
print(f"Error: {e}")
|
1238 |
-
traceback.print_exc()
|
1239 |
-
|
1240 |
-
return result
|
1241 |
-
|
1242 |
-
|
1243 |
-
def load_ms_marco_result(
|
1244 |
-
csv_result_files,
|
1245 |
-
force_recalculate=False,
|
1246 |
-
calc_bertscore=True,
|
1247 |
-
save=False,
|
1248 |
-
penalty_function="cubic",
|
1249 |
-
):
|
1250 |
-
result = {}
|
1251 |
-
for csv_result_file in csv_result_files:
|
1252 |
-
try:
|
1253 |
-
df = pd.read_csv(csv_result_file)
|
1254 |
-
model_name = get_model_name(csv_result_file)
|
1255 |
-
print(f"\tmodel_name: {model_name}")
|
1256 |
-
|
1257 |
-
dfs = [
|
1258 |
-
load_for_repetition_penalty_ms_macro(
|
1259 |
-
csv_result_file,
|
1260 |
-
repetition_penalty,
|
1261 |
-
force_recalculate=force_recalculate,
|
1262 |
-
)
|
1263 |
-
for repetition_penalty in df["repetition_penalty"]
|
1264 |
-
]
|
1265 |
-
|
1266 |
-
answer_lens = []
|
1267 |
-
for df_rpp in dfs:
|
1268 |
-
answer_lens.append(df_rpp["answer_len"].mean())
|
1269 |
-
df["answer_len"] = answer_lens
|
1270 |
-
|
1271 |
-
col = "bert_score" if calc_bertscore else "meteor"
|
1272 |
-
score_unavailable = col not in df.columns
|
1273 |
-
|
1274 |
-
if score_unavailable:
|
1275 |
-
save = True
|
1276 |
-
bert_meteor_scores = []
|
1277 |
-
bert_score_references = None
|
1278 |
-
for df_rpp in dfs:
|
1279 |
-
if calc_bertscore:
|
1280 |
-
bert_meteor_score = 0
|
1281 |
-
|
1282 |
-
for i, row in df_rpp.iterrows():
|
1283 |
-
answer = row["answer"]
|
1284 |
-
if not isinstance(answer, str):
|
1285 |
-
answer = ""
|
1286 |
-
bert_meteor_score += bert_score.compute(
|
1287 |
-
predictions=[answer],
|
1288 |
-
references=[row["ground_truth"][0]],
|
1289 |
-
lang="en",
|
1290 |
-
model_type="microsoft/deberta-large-mnli",
|
1291 |
-
)["f1"][0]
|
1292 |
-
# get average of bertscore
|
1293 |
-
bert_meteor_score = bert_meteor_score / len(df_rpp)
|
1294 |
-
|
1295 |
-
print(f"bert_score: {bert_meteor_score}")
|
1296 |
-
else:
|
1297 |
-
bert_meteor_score = meteor.compute(
|
1298 |
-
predictions=df_rpp["answer"],
|
1299 |
-
references=df_rpp["ground_truth"],
|
1300 |
-
)["meteor"]
|
1301 |
-
|
1302 |
-
bert_meteor_scores.append(bert_meteor_score)
|
1303 |
-
|
1304 |
-
df[col] = bert_meteor_scores
|
1305 |
-
|
1306 |
-
result[model_name] = {
|
1307 |
-
"df_overall": df,
|
1308 |
-
"df_list_repetition_penalty": dfs,
|
1309 |
-
"file": csv_result_file,
|
1310 |
-
}
|
1311 |
-
newline_score, repetition_score, perf, rap, nrr = calc_rap_scores(
|
1312 |
-
result[model_name],
|
1313 |
-
precision=col,
|
1314 |
-
recall=col,
|
1315 |
-
penalty_function=penalty_function,
|
1316 |
-
)
|
1317 |
-
df["newline_score"] = newline_score
|
1318 |
-
df["repetition_score"] = repetition_score
|
1319 |
-
df["total_repetitions"] = df["newline_score"] + df["repetition_score"]
|
1320 |
-
df["perf"] = perf
|
1321 |
-
df["nrr"] = nrr
|
1322 |
-
df["rap"] = rap
|
1323 |
-
df["rr"] = df["nrr"].apply(lambda x: 1 - x)
|
1324 |
-
df["rrp"] = df["rr"].apply(lambda x: x * 100)
|
1325 |
-
|
1326 |
-
if save:
|
1327 |
-
df.to_csv(csv_result_file, index=False)
|
1328 |
-
except Exception as e:
|
1329 |
-
print("An error occurred:", e)
|
1330 |
-
traceback.print_exc()
|
1331 |
-
print(f"csv_result_file: {csv_result_file}")
|
1332 |
-
|
1333 |
-
return result
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
eval_modules/utils.py
CHANGED
@@ -1,174 +1,85 @@
|
|
1 |
# -*- coding:utf-8 -*-
|
2 |
from __future__ import annotations
|
3 |
|
4 |
-
import json
|
5 |
-
import logging
|
6 |
-
import os
|
7 |
-
import platform
|
8 |
import re
|
9 |
-
from pathlib import Path
|
10 |
import evaluate
|
11 |
import pandas as pd
|
12 |
-
import requests
|
13 |
-
import torch
|
14 |
-
from tqdm import tqdm
|
15 |
-
|
16 |
-
|
17 |
-
class LogRecord(logging.LogRecord):
|
18 |
-
def getMessage(self):
|
19 |
-
msg = self.msg
|
20 |
-
if self.args:
|
21 |
-
if isinstance(self.args, dict):
|
22 |
-
msg = msg.format(**self.args)
|
23 |
-
else:
|
24 |
-
msg = msg.format(*self.args)
|
25 |
-
return msg
|
26 |
-
|
27 |
-
|
28 |
-
class Logger(logging.Logger):
|
29 |
-
def makeRecord(
|
30 |
-
self,
|
31 |
-
name,
|
32 |
-
level,
|
33 |
-
fn,
|
34 |
-
lno,
|
35 |
-
msg,
|
36 |
-
args,
|
37 |
-
exc_info,
|
38 |
-
func=None,
|
39 |
-
extra=None,
|
40 |
-
sinfo=None,
|
41 |
-
):
|
42 |
-
rv = LogRecord(name, level, fn, lno, msg, args, exc_info, func, sinfo)
|
43 |
-
if extra is not None:
|
44 |
-
for key in extra:
|
45 |
-
rv.__dict__[key] = extra[key]
|
46 |
-
return rv
|
47 |
-
|
48 |
-
|
49 |
-
def init_settings():
|
50 |
-
logging.setLoggerClass(Logger)
|
51 |
-
logging.basicConfig(
|
52 |
-
level=logging.WARNING,
|
53 |
-
format="%(asctime)s [%(levelname)s] [%(filename)s:%(lineno)d] %(message)s",
|
54 |
-
)
|
55 |
|
|
|
56 |
|
57 |
-
|
58 |
-
|
59 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
|
61 |
-
def print_llm_response(llm_response, debug_retrieval=True):
|
62 |
-
answer = llm_response["answer"] if "answer" in llm_response else None
|
63 |
-
if answer is None:
|
64 |
-
answer = llm_response["response"] if "response" in llm_response else None
|
65 |
|
66 |
-
|
67 |
-
|
68 |
-
print(answer)
|
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 |
-
"MPS not available because the current MacOS version is not 12.3+ "
|
115 |
-
"and/or you do not have an MPS-enabled device on this machine."
|
116 |
-
)
|
117 |
-
else:
|
118 |
-
device_type_available = "mps"
|
119 |
-
|
120 |
-
if torch.cuda.is_available():
|
121 |
-
print("CUDA is available, we have found ", torch.cuda.device_count(), " GPU(s)")
|
122 |
-
print(torch.cuda.get_device_name(0))
|
123 |
-
print("CUDA version: " + torch.version.cuda)
|
124 |
-
device_type_available = f"cuda:{torch.cuda.current_device()}"
|
125 |
-
|
126 |
-
return (
|
127 |
-
os.environ.get("HF_EMBEDDINGS_DEVICE_TYPE") or device_type_available,
|
128 |
-
os.environ.get("HF_PIPELINE_DEVICE_TYPE") or device_type_available,
|
129 |
)
|
|
|
|
|
|
|
|
|
130 |
|
|
|
|
|
|
|
131 |
|
132 |
-
def ensure_model_is_downloaded(llm_model_type):
|
133 |
-
if llm_model_type.startswith("gpt4all"):
|
134 |
-
local_path = (
|
135 |
-
os.environ.get("GPT4ALL_J_MODEL_PATH")
|
136 |
-
if llm_model_type == "gpt4all-j"
|
137 |
-
else os.environ.get("GPT4ALL_MODEL_PATH")
|
138 |
-
)
|
139 |
-
url = (
|
140 |
-
os.environ.get("GPT4ALL_J_DOWNLOAD_LINK")
|
141 |
-
if llm_model_type == "gpt4all-j"
|
142 |
-
else os.environ.get("GPT4ALL_DOWNLOAD_LINK")
|
143 |
-
)
|
144 |
-
elif llm_model_type == "llamacpp":
|
145 |
-
local_path = os.environ.get("LLAMACPP_MODEL_PATH")
|
146 |
-
url = os.environ.get("LLAMACPP_DOWNLOAD_LINK")
|
147 |
-
elif llm_model_type == "ctransformers":
|
148 |
-
local_path = os.environ.get("CTRANSFORMERS_MODEL_PATH")
|
149 |
-
url = os.environ.get("CTRANSFORMERS_DOWNLOAD_LINK")
|
150 |
-
else:
|
151 |
-
raise ValueError(f"wrong model typle: {llm_model_type}")
|
152 |
-
|
153 |
-
path = Path(local_path)
|
154 |
-
|
155 |
-
if path.is_file():
|
156 |
-
print(f"model: {local_path} exists")
|
157 |
-
else:
|
158 |
-
print(f"downloading model: {local_path} from {url} ...")
|
159 |
-
path.parent.mkdir(parents=True, exist_ok=True)
|
160 |
-
|
161 |
-
# send a GET request to the URL to download the file. Stream since it's large
|
162 |
-
response = requests.get(url, stream=True)
|
163 |
-
|
164 |
-
# open the file in binary mode and write the contents of the response to it in chunks
|
165 |
-
# This is a large file, so be prepared to wait.
|
166 |
-
with open(local_path, "wb") as f:
|
167 |
-
for chunk in tqdm(response.iter_content(chunk_size=8192)):
|
168 |
-
if chunk:
|
169 |
-
f.write(chunk)
|
170 |
-
|
171 |
-
return local_path
|
172 |
|
173 |
|
174 |
bleu = evaluate.load("bleu")
|
|
|
1 |
# -*- coding:utf-8 -*-
|
2 |
from __future__ import annotations
|
3 |
|
|
|
|
|
|
|
|
|
4 |
import re
|
|
|
5 |
import evaluate
|
6 |
import pandas as pd
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
+
print(f"loading: {__file__}")
|
9 |
|
10 |
+
# pattern_non_word_char_repetition = re.compile(r"\s{5,}")
|
11 |
+
# pattern_text_repetitions = re.compile(r"(.{5}.*)\s*((\1)\s*)+", re.M | re.DOTALL)
|
12 |
|
13 |
+
# final version
|
14 |
+
pattern_non_word_char_repetition = re.compile(r"[\s\W]{5,}")
|
15 |
+
pattern_text_repetitions = re.compile(
|
16 |
+
r"(?P<repeat>.{5}.*?)(?:[\s\W]*(?P=repeat))+", re.M | re.DOTALL | re.IGNORECASE
|
17 |
+
)
|
18 |
+
# Explanation of the Regex Pattern:
|
19 |
+
# (?P<repeat>.{5}.*?): Captures any sequence of characters with minimal length of 5 and names this group repeat.
|
20 |
+
# .*?: Matches zero or more characters, non-greedily (as few as possible).
|
21 |
+
# (?:[\s\W]+(?P=repeat))+: A non-capturing group that matches one or more repetitions of:
|
22 |
+
# [\s\W]+: One or more whitespace or non-word characters (spaces, punctuation, etc.).
|
23 |
+
# (?P=repeat): A backreference to the named group repeat.
|
24 |
|
|
|
|
|
|
|
|
|
25 |
|
26 |
+
def del_non_word_char_repetition(text, debug=False):
|
27 |
+
count = 0
|
|
|
28 |
|
29 |
+
if isinstance(text, str):
|
30 |
+
if debug:
|
31 |
+
print("----detect non-word characters repetition----")
|
32 |
+
count = len(text)
|
33 |
+
text = pattern_non_word_char_repetition.sub("\t", text)
|
34 |
+
count -= len(text)
|
35 |
+
if debug and count:
|
36 |
+
print(f"removed non-word characters repetition: {count}")
|
37 |
+
return text, count
|
38 |
+
|
39 |
+
|
40 |
+
# final version for repetition detection
|
41 |
+
def detect_text_repetitions(text, debug=False):
|
42 |
+
count = 0
|
43 |
+
|
44 |
+
if isinstance(text, str):
|
45 |
+
if debug:
|
46 |
+
print("----detect text repetitions----")
|
47 |
+
matches = pattern_text_repetitions.finditer(text)
|
48 |
+
for match in matches:
|
49 |
+
if debug:
|
50 |
+
print(match)
|
51 |
+
for groupNum in range(0, len(match.groups())):
|
52 |
+
groupNum = groupNum + 1
|
53 |
+
print(
|
54 |
+
"Group {groupNum} found at {start}-{end}: `{group}`".format(
|
55 |
+
groupNum=groupNum,
|
56 |
+
start=match.start(groupNum),
|
57 |
+
end=match.end(groupNum),
|
58 |
+
group=match.group(groupNum),
|
59 |
+
)
|
60 |
+
)
|
61 |
+
|
62 |
+
start, end = match.span()
|
63 |
+
count += end - start - len(match.group(1))
|
64 |
+
|
65 |
+
return count
|
66 |
+
|
67 |
+
|
68 |
+
def detect_repetitions(text, debug=False):
|
69 |
+
if isinstance(text, str) is False:
|
70 |
+
return 0, 0, 0
|
71 |
+
text, count_non_word_char_repetition = del_non_word_char_repetition(
|
72 |
+
text, debug=debug
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
)
|
74 |
+
count_text_repetitions = detect_text_repetitions(text, debug=debug)
|
75 |
+
total_repetitions = count_non_word_char_repetition + count_text_repetitions
|
76 |
+
|
77 |
+
result = (count_non_word_char_repetition, count_text_repetitions, total_repetitions)
|
78 |
|
79 |
+
if debug:
|
80 |
+
print(result)
|
81 |
+
return result
|
82 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
83 |
|
84 |
|
85 |
bleu = evaluate.load("bleu")
|