Update README.md

README.md

@@ -18,6 +18,18 @@ license: apache-2.0
 - Adapted LLaMA version: [MiniCPM-S-1B-sft-llama-format](https://huggingface.co/openbmb/MiniCPM-S-1B-sft-llama-format/)
 - Adapted PowerInfer version: [MiniCPM-S-1B-sft-gguf](https://huggingface.co/openbmb/MiniCPM-S-1B-sft-gguf)
 
+### Chat Template
+
+To make the model sophisticatedly respond to a query, it is recommended to use a standard chat prompt, such as:
+
+```
+<用户>{prompt}<AI>
+```
+
+where `prompt` is the query text, while `<用户>` and `<AI>` are prompt tokens.
+
+Also, make sure that you have **a bos token `<s>` at the beginning of any input**, or the model can sometimes behave improperly.
+
 ### Introduction
 
 The utilization of activation sparsity, namely the existence of considerable weakly-contributed elements among activation outputs, is a promising method for inference acceleration of large language models (LLMs) ([Liu et al., 2023](https://proceedings.mlr.press/v202/liu23am/liu23am.pdf); [Song et al., 2023](https://arxiv.org/pdf/2312.12456.pdf)). Concretely, acceleration methods based on activation sparsity usually achieve higher inference speed by making wiser resource allocation and computation policies to avoid resource waste on these weakly-contributed parameters.