Scaling Instruction-Finetuned Language Models
擴充套件指令 - 微調語言模型
What's new?
有什麼新的？
Image Source: Scaling Instruction-Finetuned Language Models (opens in a new tab)
圖片來源：Scaling Instruction-Finetuned Language Models (opens in a new tab)
This paper explores the benefits scaling instruction finetuning (opens in a new tab) and how it improves performance on a variety of models (PaLM, T5), prompting setups (zero-shot, few-shot, CoT), and benchmarks (MMLU, TyDiQA). This is explored with the following aspects: scaling the number of tasks (1.8K tasks), scaling model size, and finetuning on chain-of-thought data (9 datasets used).
本文探討了擴充套件指令微調的好處，以及它如何提高各種模型（PaLM、T5）、提示設定（zero-shot、few-shot、CoT）和基準測試（MMLU、TyDiQA）的效能。這是透過以下方面進行探討的：擴充套件任務數量（1.8K個任務）、擴充套件模型大小，以及在思維鏈資料上進行微調（使用了9個數據集）。
Finetuning procedure:
Finetuning程式：
- 1.8K tasks were phrased as instructions and used to finetune the model
- Uses both with and without exemplars, and with and without CoT
- 1.8K個任務被表述為指令，用於微調模型
- 同時使用有和沒有範例，以及有和沒有CoT
Finetuning tasks and held out tasks shown below:
Finetuning 任務和保留任務如下所示：
Capabilities & Key Results
能力和關鍵結果
- Instruction finetuning scales well with the number of tasks and the size of the model; this suggests the need for scaling number of tasks and size of model further
- Adding CoT datasets into the finetuning enables good performance on reasoning tasks
- Flan-PaLM has improved multilingual abilities; 14.9% improvement on one-shot TyDiQA; 8.1% improvement on arithmetic reasoning in under-represented languages
- Plan-PaLM also performs well on open-ended generation questions, which is a good indicator for improved usability
- Improves performance across responsible AI (RAI) benchmarks
- Flan-T5 instruction tuned models demonstrate strong few-shot capabilities and outperforms public checkpoint such as T5
- 指令微調隨著任務數量和模型大小的增加而擴充套件得很好；這表明需要進一步擴充套件任務數量和模型大小
- 將CoT資料集新增到微調中可以在推理任務上實現良好的效能
- Flan-PaLM具有改進的多語言能力；在一次性TyDiQA上提高了14.9％；在代表性不足的語言中進行算術推理的提高了8.1％
- Plan-PaLM在開放式產生問題上也表現良好，這是改進可用性的良好指標
- 提高了負責任的AI（RAI）基準的效能
- Flan-T5指令微調模型展示了強大的少樣本能力，並且優於T5等公共檢查點
The results when scaling number of finetuning tasks and model size: scaling both the size of the model and the number of finetuning tasks is expected to continue improving performance, although scaling the number of tasks has diminished returns.
當調整微調任務數量和模型大小時的結果： 預期同時調整模型大小和微調任務數量將繼續提高效能，但調整任務數量的效益會逐漸減少。
Image Source: Scaling Instruction-Finetuned Language Models (opens in a new tab)
圖片來源：Scaling Instruction-Finetuned Language Models (opens in a new tab)
The results when finetuning with non-CoT and CoT data: Jointly finetuning on non-CoT and CoT data improves performance on both evaluations, compared to finetuning on just one or the other.
**使用非 CoT 和 CoT 資料進行微調的結果：**與僅在其中一個資料上進行微調相比，同時在非 CoT 和 CoT 資料上進行微調可以提高兩個評估的效能。
Image Source: Scaling Instruction-Finetuned Language Models (opens in a new tab)
圖片來源：Scaling Instruction-Finetuned Language Models (opens in a new tab)
In addition, self-consistency combined with CoT achieves SoTA results on several benchmarks. CoT + self-consistency also significantly improves results on benchmarks involving math problems (e.g., MGSM, GSM8K).
此外，自我一致性結合CoT在多個基準測試中實現了SoTA結果。CoT + 自我一致性還顯著改善了涉及數學問題的基準測試結果（例如，MGSM，GSM8K）。
Image Source: Scaling Instruction-Finetuned Language Models (opens in a new tab)
圖片來源：Scaling Instruction-Finetuned Language Models (opens in a new tab)
CoT finetuning unlocks zero-shot reasoning, activated by the phrase "let's think step-by-step", on BIG-Bench tasks. In general, zero-shot CoT Flan-PaLM outperforms zero-shot CoT PaLM without finetuning.
CoT 微調可以解鎖zero-shot 推理，透過“讓我們逐步思考”短語在 BIG-Bench 任務中啟動。一般來說，zero-shot CoT Flan-PaLM 優於沒有微調的zero-shot CoT PaLM。
Image Source: Scaling Instruction-Finetuned Language Models (opens in a new tab)
圖片來源：Scaling Instruction-Finetuned Language Models (opens in a new tab)
Below are some demonstrations of zero-shot CoT for PaLM and Flan-PaLM in unseen tasks.
以下是 PaLM 和 Flan-PaLM 在未知任務中的零樣本 CoT 示範。
Image Source: Scaling Instruction-Finetuned Language Models (opens in a new tab)
圖片來源：Scaling Instruction-Finetuned Language Models (opens in a new tab)
Below are more examples for zero-shot prompting. It shows how the PaLM model struggles with repetitions and not replying to instructions in the zero-shot setting where the Flan-PaLM is able to perform well. Few-shot exemplars can mitigate these errors.
以下是更多的zero-shot提示示範。它展示了 PaLM 模型在重複和不回覆zero-shot設定中的指令方面的困難，而 Flan-PaLM 能夠表現良好。少量示範可以緩解這些錯誤。
Image Source: Scaling Instruction-Finetuned Language Models (opens in a new tab)
圖片來源：Scaling Instruction-Finetuned Language Models (opens in a new tab)
Below are some examples demonstrating more zero-shot capabilities of the Flan-PALM model on several different types of challenging open-ended questions:
以下是一些示範，展示了Flan-PALM模型在幾種不同型別的具有挑戰性的開放式問題上的更多zero-shot能力：
Image Source: Scaling Instruction-Finetuned Language Models (opens in a new tab)
圖片來源：Scaling Instruction-Finetuned Language Models (opens in a new tab)
Image Source: Scaling Instruction-Finetuned Language Models (opens in a new tab)
圖片來源：Scaling Instruction-Finetuned Language Models (opens in a new tab)
Image Source: Scaling Instruction-Finetuned Language Models (opens in a new tab)
圖片來源：Scaling Instruction-Finetuned Language Models (opens in a new tab)
You can try Flan-T5 models on the Hugging Face Hub (opens in a new tab).
你可以在 Hugging Face Hub 上嘗試 Flan-T5 模型 (opens in a new tab)。