File size: 8,446 Bytes
da6e1bc 731eddd da6e1bc 731eddd da6e1bc ce2acb0 eaf2d97 da6e1bc 731eddd da6e1bc 731eddd da6e1bc 731eddd da6e1bc 731eddd da6e1bc 731eddd da6e1bc 731eddd da6e1bc 731eddd da6e1bc 731eddd da6e1bc ce2acb0 731eddd da6e1bc 731eddd da6e1bc 8274634 ce2acb0 da6e1bc 731eddd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 |
import random
from functools import partial
import evaluate
import pandas as pd
import sentencepiece as spm
from datasets_.flores import flores_sentences
from joblib.memory import Memory
from languages import languages, script_name
from models import complete, transcribe
from datasets import load_dataset
cache = Memory(location=".cache", verbose=0).cache
bleu = evaluate.load("bleu")
chrf = evaluate.load("chrf")
wer = evaluate.load("wer")
tokenizer = spm.SentencePieceProcessor(
model_file="data/spbleu/flores200_sacrebleu_tokenizer_spm.model"
)
# sample languages to translate to
target_languages = languages[languages["in_benchmark"]].sample(
frac=1, weights="speakers", replace=True, random_state=42
)
@cache
async def translate_and_evaluate(model, bcp_47, sentence_nr, mode="from"):
original_language = languages[languages["bcp_47"] == bcp_47].iloc[0]
target_language = target_languages.iloc[sentence_nr]
match mode:
case "from":
pass
case "to":
original_language, target_language = target_language, original_language
original_sentence = flores_sentences(original_language)[sentence_nr].strip()
target_sentence = flores_sentences(target_language)[sentence_nr].strip()
script = script_name(target_language.flores_path.split("_")[1])
reply = await complete(
model=model,
messages=[
{
"role": "user",
"content": f"Translate the following text to the {target_language.language_name} language; use the {script} script; reply only with the translation:\n\n{original_sentence}",
}
],
temperature=0,
max_tokens=1024,
)
prediction = reply.choices[0].message.content.strip()
if prediction.strip():
bleu_score = bleu.compute(
predictions=[prediction],
references=[target_sentence],
tokenizer=tokenizer.tokenize,
)
else:
bleu_score = {"bleu": 0}
chrf_score = chrf.compute(predictions=[prediction], references=[target_sentence])
return [
{
"model": model,
"bcp_47": bcp_47,
"task": f"translation_{mode}",
"metric": metric,
"score": score,
"sentence_nr": sentence_nr,
}
for metric, score in (
("bleu", bleu_score["bleu"]),
("chrf", chrf_score["score"] / 100),
)
]
metadata = pd.read_csv("data/floresp-v2.0-rc.3/metadata_dev.tsv", sep="\t")
@cache
async def classify_and_evaluate(model, bcp_47, nr):
language = languages[languages["bcp_47"] == bcp_47].iloc[0]
sentences = pd.DataFrame(flores_sentences(language), columns=["text"])
sentences = pd.concat([metadata, sentences], axis=1)
sentences = sentences.dropna(subset=["topic"])
sentences["topic"] = sentences["topic"].str.lower()
paragraphs = (
sentences.groupby("URL").agg({"text": " ".join, "topic": "first"}).reset_index()
)
top_topics = paragraphs.value_counts("topic").head(5).index
paragraphs = paragraphs[paragraphs["topic"].isin(top_topics)]
examples = pd.concat(
[
paragraphs[paragraphs["topic"] == t].sample(n=5, random_state=42)
for t in top_topics
]
).sample(frac=1, random_state=42)
test_paragraphs = paragraphs[~paragraphs["URL"].isin(examples["URL"])].sample(
frac=1, random_state=42
)
test_paragraph = test_paragraphs.iloc[nr]
def topic_to_number(topic):
return top_topics.get_loc(topic)
messages = []
for example in examples.itertuples():
messages += [
{"role": "user", "content": example.text},
{"role": "assistant", "content": str(topic_to_number(example.topic))},
]
reply = await complete(
model=model,
messages=[
*messages,
{
"role": "user",
"content": test_paragraph.text,
},
],
temperature=0,
max_tokens=5,
)
try:
pred = int(reply.choices[0].message.content.strip())
except ValueError:
pred = -1
true = topic_to_number(test_paragraph.topic)
return [
{
"model": model,
"bcp_47": bcp_47,
"task": "classification",
"metric": "accuracy",
"score": int(pred == true),
"sentence_nr": nr,
}
]
def corrupt_sentence(sentence):
# replace 5% of the sentence with <mask>
mask_length = round(len(sentence) * 0.05)
start = random.randint(0, len(sentence) - mask_length)
end = start + mask_length
return sentence[:start] + "<mask>" + sentence[end:]
@cache
async def mlm_and_evaluate(model, language_bcp_47, nr):
language = languages[languages["bcp_47"] == language_bcp_47].iloc[0]
sentences = pd.DataFrame(flores_sentences(language), columns=["text"])
sentences["corrupt_text"] = sentences["text"].apply(corrupt_sentence)
examples = sentences.sample(n=10, random_state=42)
test_sentences = sentences[~sentences["text"].isin(examples["text"])].sample(
frac=1, random_state=42
)
test_sentence = test_sentences.iloc[nr]
messages = []
for example in examples.itertuples():
messages += [
{"role": "user", "content": example.corrupt_text},
{"role": "assistant", "content": example.text},
]
reply = await complete(
model=model,
messages=[
*messages,
{
"role": "user",
"content": test_sentence.corrupt_text,
},
],
temperature=0,
max_tokens=1024,
)
prediction = reply.choices[0].message.content.strip()
chrf_score = chrf.compute(predictions=[prediction], references=[test_sentence.text])
return [
{
"model": model,
"bcp_47": language["bcp_47"],
"task": "language_modeling",
"metric": "chrf",
"score": chrf_score["score"] / 100,
"sentence_nr": nr,
}
]
@cache
def _load_dataset(dataset, subset):
return load_dataset(dataset, subset)
@cache
async def mmlu_and_evaluate(model, language_bcp_47, nr):
data = _load_dataset("CohereForAI/Global-MMLU", language_bcp_47)
item = data["test"][nr]
def format_item(item):
return f"""{item['question']}
A: {item['option_a']}
B: {item['option_b']}
C: {item['option_c']}
D: {item['option_d']}
A|B|C|D?"""
messages = []
for example in data["dev"].select(range(5)):
messages += [{"role": "user", "content": format_item(example)}, {"role": "assistant", "content": example["answer"]}]
messages += [{"role": "user", "content": format_item(item)}]
reply = await complete(
model=model,
messages=messages,
temperature=0,
max_tokens=1,
)
print(reply.choices[0].message.content.strip())
acc = int(reply.choices[0].message.content.strip() == item["answer"])
return [
{
"model": model,
"bcp_47": language_bcp_47,
"task": "mmlu",
"metric": "accuracy",
"score": acc,
"sentence_nr": nr,
}
]
@cache
async def transcribe_and_evaluate(model, language_bcp_47, nr):
language = languages[languages["bcp_47"] == language_bcp_47].iloc[0]
fleurs = pd.read_csv(
f"data/fleurs/{language.fleurs_tag}/dev.tsv",
sep="\t",
names=[
"id",
"fname",
"raw_transcription",
"transcription",
"words",
"id2",
"gender",
],
)
item = fleurs.iloc[nr]
path = f"data/fleurs/{language.fleurs_tag}/audio/dev/{item.fname}"
pred = await transcribe(path, model=model)
wer_score = wer.compute(predictions=[pred], references=[item.transcription])
return [
{
"model": model,
"bcp_47": language["bcp_47"],
"task": "asr",
"metric": "wer",
"score": wer_score,
"sentence_nr": nr,
}
]
tasks = [
partial(translate_and_evaluate, mode="from"),
partial(translate_and_evaluate, mode="to"),
classify_and_evaluate,
# mlm_and_evaluate,
mmlu_and_evaluate,
# transcribe_and_evaluate,
]
|