[TOC]
一面
项目
多模态评论生成的指标，人工准确率是多少
mplug的结构，纯LLM效果怎么样
预训练细节
手撕
输入文件1: 1kw行title，中文，1k
输入文件2 : 10w 个实体词，中文，最大100
输出：统计文件1中，出现频次最高的100个实体词（来自文件2）及其频次
# encoding: utf-8
file1 = []
file2 = []
ma = DefaultDict()
vocab_size = 10000
tokenid = {c:i for i,c in enumerate(vocab)}
class Trie:
def __init__(self):
self.child = [None]*vocab_size
self.isEnd = False
def insert(self, s):
root = self
for c in s:
index = tokenid[c]
if not root.child[index]:
root.child[index] = Trie()
root = root.child[index]
root.isEnd = True
def find(self, s):
root = self
for c in s:
index = tokenid[c]
if not root.child[index]:
return False
root = root.child[index]
return root.isEnd
def count(self,s):
root = self
for i,c in enumerate(s):
index = tokenid[c]
if not root.child[index]:
return
root = root.child[index]
if root.isEnd:
ma[s[:i+1]]+=1
tree = tire()
for word in file2:
tree.insert(word)
for title in file1:
n = len(title)
for i in range(n):
tree.count(title[i:])
# 选取可以用快选
heap = []
for word,fre in ma:
heapq.heappush(heap,(-fre,word))
res = []
for i in range(100):
fre,word = heapq.heappop(heap)
res.append((word,-fre))
print(res)
二面
长文本问题？transformer什么部分是长文本的bottleneck？
pretrain相关技术，rlhf、ppo相关的文章读过吗？
一个开放问题，现在有100条优质创作数据，我们大概有一个70B模型，怎么从模型本身提升模型创作能力？
手撕：判断一个字符串能否被给定字典中的字符串划分
s = "catsandog"
wordDict = ["cats", "dog", "sand", "and", "cat"]
class Node:
def __init__(self) -> None:
self.children = [None]*26
self.is_end = False
class Tire:
def __init__(self) -> None:
self.root = Node()
def insert(self,word):
root = self.root
for c in word:
if not root.children[ord(c)-ord('a')]:
root.children[ord(c)-ord('a')]=Node()
root = root.children[ord(c)-ord('a')]
root.is_end = True
t = Tire()
for word in wordDict:
t.insert(word)
n = len(s)
b = [False]*(n+1)
b[-1]=True
for i in range(n-1,-1,-1):
root = t.root
for j in range(i,n):
c = s[j]
if not root.children[ord(c)-ord('a')]:
break
root = root.children[ord(c)-ord('a')]
if root.is_end:
b[i] = b[j+1] if b[j+1] else b[i]
print("true" if b[0] else "false")
