---
license: mit
task_categories:
- question-answering
language:
- en
tags:
- finance
size_categories:
- 1K<n<10K
---
This data comprises synthetic question and answer pairs created by GPT-4-turbo on SEC filings for 29 companies. The dataset has the following columns:
questions, answers, chunks and sorted_chunks.
questions: the list of questions, there were 5 questions created for a 2000 word section of different SEC filings.
answers: the answer generated by GPT-4.
chunks: these are the bits of text that are segmented.
sorted_chunks: these are the chunks being sorted, using Dense Passage Retrieval (DPR).
The sorting is done by using the embeddings extracted from the BERT models fine-tuned for query and context embedding for DPR. The sorting is then done by using FAISS using IndexFlatL2.
The following is an example of how one can sort the chunks using DPR pre-trained embeddings and FAISS:
First the DPR models should be imported as such:
```python
from transformers import DPRConfig, DPRContextEncoder, DPRQuestionEncoder, DPRQuestionEncoderTokenizer, DPRContextEncoderTokenizer
# context tokenizer and encoder
tokenizer_c = DPRContextEncoderTokenizer.from_pretrained("facebook/dpr-ctx_encoder-single-nq-base")
model_c = DPRContextEncoder.from_pretrained("facebook/dpr-ctx_encoder-single-nq-base", device_map = 'cuda')
# question tokenizer and encoder
tokenizer_q = DPRQuestionEncoderTokenizer.from_pretrained("facebook/dpr-question_encoder-single-nq-base")
model_q = DPRQuestionEncoder.from_pretrained("facebook/dpr-question_encoder-single-nq-base", device_map = 'cuda')
```
The created chunks can then be sorted using FAISS as such:
```python
import faiss
sorted_chunks = []
print(f'----- total rows: {len(answers)}\n')
for q,a,c in tqdm(zip(questions[:], answers[:], chunks[:])):
index = faiss.IndexFlatL2(768) # Assuming embeddings are of size 768
# Encode texts using context encoder
encoded_texts = tokenizer_c(c, return_tensors='pt', padding = 'max_length')['input_ids']
encoded_texts = model_c(encoded_texts.cuda()).pooler_output
# Encode question using question encoder
encoded_question = tokenizer_q(q, return_tensors='pt')['input_ids']
encoded_question = model_q(encoded_question.cuda()).pooler_output
if len(c)>4:
# Perform FAISS search to get top 5 texts
faiss_texts = encoded_texts.detach().cpu().numpy()
faiss_question = encoded_question.detach().cpu().numpy()
# faiss_question = encoded_question.detach().cpu().numpy()
index.add(np.array(faiss_texts))
_, faiss_topk = index.search(faiss_question, k=4)
top_texts = [c[j] for j in faiss_topk[0]]
top_t = ''
for j in top_texts:
top_t += j + ', '
sorted_chunks.append(top_t)
else:
# Perform FAISS search to get top 5 texts
faiss_texts = encoded_texts.detach().cpu().numpy()
faiss_question = encoded_question.detach().cpu().numpy()
# faiss_question = encoded_question.detach().cpu().numpy()
index.add(np.array(faiss_texts))
_, faiss_topk = index.search(faiss_question, k=len(c))
top_texts = [c[j] for j in faiss_topk[0]]
top_t = ''
for j in top_texts:
top_t += j + ', '
sorted_chunks.append(top_t)
df = pd.DataFrame(data = {'questions':questions[:],
'answers':answers[:], 'chunks':chunks[:], 'sorted_chunks':sorted_chunks})
```
### 数据集元信息
许可证：MIT许可证
任务类别：问答
语言：英语
标签：金融
规模类别：1000 < 样本量 < 10000
本数据集包含由GPT-4-turbo基于29家公司的美国证券交易委员会（U.S. Securities and Exchange Commission, SEC）备案文件生成的合成问答对。数据集包含以下四列：
1. `questions`：问题列表，针对不同SEC备案文件中每2000词的文本片段，会生成5个对应问题；
2. `answers`：由GPT-4生成的答案内容；
3. `chunks`：经过分段处理的文本片段；
4. `sorted_chunks`：通过密集段落检索（Dense Passage Retrieval, DPR）完成相关性排序的文本片段。
该排序流程依托针对DPR的查询与上下文嵌入任务微调的BERT（Bidirectional Encoder Representations from Transformers）模型提取的嵌入向量实现，随后借助FAISS的IndexFlatL2索引完成排序操作。
以下展示了如何使用DPR预训练嵌入与FAISS实现文本片段排序：
python
from transformers import DPRConfig, DPRContextEncoder, DPRQuestionEncoder, DPRQuestionEncoderTokenizer, DPRContextEncoderTokenizer
# 上下文分词器与编码器
tokenizer_c = DPRContextEncoderTokenizer.from_pretrained("facebook/dpr-ctx_encoder-single-nq-base")
model_c = DPRContextEncoder.from_pretrained("facebook/dpr-ctx_encoder-single-nq-base", device_map = 'cuda')
# 问题分词器与编码器
tokenizer_q = DPRQuestionEncoderTokenizer.from_pretrained("facebook/dpr-question_encoder-single-nq-base")
model_q = DPRQuestionEncoder.from_pretrained("facebook/dpr-question_encoder-single-nq-base", device_map = 'cuda')
随后可通过FAISS对生成的文本片段进行排序：
python
import faiss
sorted_chunks = []
print(f'----- 总记录数: {len(answers)}
')
for q,a,c in tqdm(zip(questions[:], answers[:], chunks[:])):
index = faiss.IndexFlatL2(768) # 假设嵌入向量维度为768
# 使用上下文编码器对文本进行编码
encoded_texts = tokenizer_c(c, return_tensors='pt', padding = 'max_length')['input_ids']
encoded_texts = model_c(encoded_texts.cuda()).pooler_output
# 使用问题编码器对查询问题进行编码
encoded_question = tokenizer_q(q, return_tensors='pt')['input_ids']
encoded_question = model_q(encoded_question.cuda()).pooler_output
if len(c)>4:
# 执行FAISS搜索以获取前5个相关文本片段
faiss_texts = encoded_texts.detach().cpu().numpy()
faiss_question = encoded_question.detach().cpu().numpy()
# faiss_question = encoded_question.detach().cpu().numpy()
index.add(np.array(faiss_texts))
_, faiss_topk = index.search(faiss_question, k=4)
top_texts = [c[j] for j in faiss_topk[0]]
top_t = ''
for j in top_texts:
top_t += j + ', '
sorted_chunks.append(top_t)
else:
# 执行FAISS搜索以获取全部相关文本片段
faiss_texts = encoded_texts.detach().cpu().numpy()
faiss_question = encoded_question.detach().cpu().numpy()
# faiss_question = encoded_question.detach().cpu().numpy()
index.add(np.array(faiss_texts))
_, faiss_topk = index.search(faiss_question, k=len(c))
top_texts = [c[j] for j in faiss_topk[0]]
top_t = ''
for j in top_texts:
top_t += j + ', '
sorted_chunks.append(top_t)
df = pd.DataFrame(data = {'questions':questions[:],
'answers':answers[:], 'chunks':chunks[:], 'sorted_chunks':sorted_chunks})
