引導式測驗 · 硬 等級
BloombergGPT 與金融領域大型語言模型 (LLM) 測驗
了解 BloombergGPT 的研究設置,比較財務模型和一般模型,並根據過時的來源證據評估財務產出。
問題 1 的 8
What did the BloombergGPT paper report about its pretraining data?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- What did the BloombergGPT paper report about its pretraining data?
- A mixture of financial and general-purpose text
- Only private analyst recommendations
- Only unlabeled stock-price time series
- A corpus consisting entirely of SEC filings
- What does a strong score on one financial benchmark establish?
- The model is safe for every trading decision
- Performance on that benchmark under its test setup
- That every financial fact is current
- That a model will outperform on all languages
- Why evaluate numerical extraction separately from fluent summaries?
- Fluency guarantees arithmetic
- Numbers do not appear in financial documents
- A polished summary can still contain an incorrect figure or period
- Numerical tests cannot use source documents
- How should a retrieved filing be used in a financial LLM workflow?
- Assume the model will calculate every figure correctly
- Remove the filing date to simplify the prompt
- Treat a citation as investment advice
- Verify claims and calculations against the dated filing
- Which evaluation set is useful when comparing general and finance-specialized models?
- The same representative held-out examples for both models
- Different questions chosen to favor each model
- Only prompts used during training
- A set containing only easy positive cases
- Why check a financial model for temporal leakage?
- To make its output more persuasive
- Future information in a test example can make results look unrealistically strong
- To remove all source documents
- It is required to increase model size
- What should a model do when asked about a figure in a filing?
- Invent a likely number when retrieval fails
- Hide the currency and period
- Preserve units and reporting period and support the figure with the source
- Use a recent figure from another company
- How should teams use BloombergGPT’s published benchmark findings?
- As a guarantee for every financial task
- As proof proprietary data always improve models
- As a replacement for their own validation
- As evidence limited to the benchmarks and methods reported