David Pomerenke
commited on
Commit
·
0384b92
1
Parent(s):
b0c61ed
Shorter classification prompt + error handling
Browse files- evals/tasks.py +34 -23
evals/tasks.py
CHANGED
@@ -90,48 +90,59 @@ async def classify_and_evaluate(model, bcp_47, nr):
|
|
90 |
paragraphs = paragraphs[paragraphs["topic"].isin(top_topics)]
|
91 |
examples = pd.concat(
|
92 |
[
|
93 |
-
paragraphs[paragraphs["topic"] == t].sample(n=
|
94 |
for t in top_topics
|
95 |
]
|
96 |
-
).sample(frac=1, random_state=
|
97 |
test_paragraphs = paragraphs[~paragraphs["URL"].isin(examples["URL"])].sample(
|
98 |
frac=1, random_state=42
|
99 |
)
|
100 |
test_paragraph = test_paragraphs.iloc[nr]
|
101 |
|
102 |
-
def
|
103 |
-
return
|
104 |
|
105 |
messages = []
|
106 |
for example in examples.itertuples():
|
107 |
messages += [
|
108 |
-
{"role": "user", "content": example.text},
|
109 |
-
{"role": "assistant", "content":
|
110 |
]
|
111 |
-
|
112 |
-
|
113 |
-
messages=[
|
114 |
-
*messages,
|
115 |
-
{
|
116 |
-
"role": "user",
|
117 |
-
"content": test_paragraph.text,
|
118 |
-
},
|
119 |
-
],
|
120 |
-
temperature=0,
|
121 |
-
max_tokens=5,
|
122 |
-
)
|
123 |
try:
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
128 |
return [
|
129 |
{
|
130 |
"model": model,
|
131 |
"bcp_47": bcp_47,
|
132 |
"task": "classification",
|
133 |
"metric": "accuracy",
|
134 |
-
"score":
|
135 |
"sentence_nr": nr,
|
136 |
}
|
137 |
]
|
|
|
90 |
paragraphs = paragraphs[paragraphs["topic"].isin(top_topics)]
|
91 |
examples = pd.concat(
|
92 |
[
|
93 |
+
paragraphs[paragraphs["topic"] == t].sample(n=1, random_state=42)
|
94 |
for t in top_topics
|
95 |
]
|
96 |
+
).sample(frac=1, random_state=nr)
|
97 |
test_paragraphs = paragraphs[~paragraphs["URL"].isin(examples["URL"])].sample(
|
98 |
frac=1, random_state=42
|
99 |
)
|
100 |
test_paragraph = test_paragraphs.iloc[nr]
|
101 |
|
102 |
+
def format_prompt(text):
|
103 |
+
return f"{text}\n\nTopic: {'|'.join(top_topics)}?"
|
104 |
|
105 |
messages = []
|
106 |
for example in examples.itertuples():
|
107 |
messages += [
|
108 |
+
{"role": "user", "content": format_prompt(example.text)},
|
109 |
+
{"role": "assistant", "content": example.topic},
|
110 |
]
|
111 |
+
# some models have poor tokenization for some languages, and the prompt for this task is relatively long, so it sometimes exceeds the context window
|
112 |
+
# this is not just to blame on the context window but mostly on the model's tokenization, so we assign 0 accuracy in this case
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
try:
|
114 |
+
reply = await complete(
|
115 |
+
model=model,
|
116 |
+
messages=[
|
117 |
+
*messages,
|
118 |
+
{
|
119 |
+
"role": "user",
|
120 |
+
"content": format_prompt(test_paragraph.text),
|
121 |
+
},
|
122 |
+
],
|
123 |
+
temperature=0,
|
124 |
+
max_tokens=30,
|
125 |
+
)
|
126 |
+
response = reply.choices[0].message.content.strip().lower()
|
127 |
+
true = test_paragraph.topic
|
128 |
+
others = [t for t in top_topics if t != true]
|
129 |
+
acc = int(
|
130 |
+
response.startswith(true)
|
131 |
+
or (true in response and not any(o in response for o in others))
|
132 |
+
)
|
133 |
+
except Exception as e:
|
134 |
+
if "`inputs` tokens + `max_new_tokens` must be <= 4097" in str(e):
|
135 |
+
print(f"Max tokens exceeded for {model} in {bcp_47}")
|
136 |
+
acc = 0
|
137 |
+
else:
|
138 |
+
raise e
|
139 |
return [
|
140 |
{
|
141 |
"model": model,
|
142 |
"bcp_47": bcp_47,
|
143 |
"task": "classification",
|
144 |
"metric": "accuracy",
|
145 |
+
"score": acc,
|
146 |
"sentence_nr": nr,
|
147 |
}
|
148 |
]
|