Merge pull request #2545 from AI-Hypercomputer:sft_eval

PiperOrigin-RevId: 826614476

Author: Google-ML-Automation

src/MaxText/configs/base.yml

@@ -489,6 +489,7 @@ tokenizer_path: "src/MaxText/assets/tokenizer.llama2"
 # hf pipeline only supports huggingface type, and will ignore tokenizer_type flag
 tokenizer_type: "sentencepiece" # Currently supporting: "tiktoken", "sentencepiece", "huggingface"
 use_chat_template: False
+chat_template_path: "" # path to chat template json file
 tokenize_train_data: True  # False if the dataset is pre-tokenized
 tokenize_eval_data: True  # False if the dataset is pre-tokenized
 add_bos: True

src/MaxText/examples/chat_templates/math_qa.json

@@ -0,0 +1,5 @@
+{
+  "PROMPT_TEMPLATE": "You are given a mathematical problem. You must solve the problem and provide your reasoning. Place your entire thought process and steps between <reasoning> and </reasoning>. After your reasoning, provide only the final numerical answer, extracted from your reasoning steps, between <answer> and </answer>. The user's problem is:\n{question}",
+  "COMPLETION_TEMPLATE": "<reasoning>\n{reasoning}\n</reasoning>\n<answer>{answer}</answer>",
+  "REASONING_ANSWER_SEPARATOR": "####"
+}