enable conv history

app.py

@@ -27,7 +27,7 @@ LICENSE = """
 <p/>
 
 ---
-**A chatbot built with ReFT and Llama-2 7B. It is trained with 1K training examples from the unpaired [Ultrafeedback dataset](https://huggingface.co/datasets/openbmb/UltraFeedback).** You can train your own ReFT agent and share it on HuggingFace by following this [tutorial](https://github.com/stanfordnlp/pyreft/tree/main/examples/gradio/train_and_share.ipynb)!
+**A chatbot built with ReFT and Llama-2 7B. It is trained with 1K training examples from the unpaired [Ultrafeedback dataset](https://huggingface.co/datasets/openbmb/UltraFeedback).** It is not good at multi-turn conversations. You can train your own ReFT agent and share it on HuggingFace by following this [tutorial](https://github.com/stanfordnlp/pyreft/tree/main/examples/gradio/train_and_share.ipynb)!
 
 ---
 **This should only be used for research purposes.** We did not conduct additional safety training with ReFT. We evaluate this model using [Alpaca-eval](https://github.com/tatsu-lab/alpaca_eval). Performance results can be found in [our ReFT paper](https://arxiv.org/abs/2404.03592). Our model inherits all the underlying risks associated with Llama. See terms outlined below.
@@ -78,7 +78,7 @@ def generate(
         conversation += [f"user: {user} assistant : {assistant}"]
     conversation += [message]
     conversation = "\n".join(conversation)
-    prompt = prompt_no_input_template % message
+    prompt = prompt_no_input_template % conversation
     prompt = tokenizer(prompt, return_tensors="pt").to(model.device)
     input_ids = prompt["input_ids"]
     attention_mask = prompt["attention_mask"]