tokenizer.batch_decode解码结果为何不是字符串？

[shuxjweb]

rollout()函数中调用tokenizer.batch_decode后得到的completions为何不是字符串列表，而是token IDs？
sequence_ids = model.generate(**model_inputs, generation_config=generation_config) # [8, 1024]
completions = tokenizer.batch_decode(
sequence_ids[:, input_ids.shape[1]:], skip_special_tokens=True
)

跑了一晚上，解码的结果依然如下所示的token列表，而非字符串：
'<|SPEECH_GENERATION_START|><|s_64943|><|s_33339|><|s_63590|>,...'

但接下来计算reward时又是通过字符串来匹配，导致answer_match一直是None，reward一直是0：
answer_match = re.search(
r"(.*?)",
completion,
flags=re.DOTALL,
)

训很久后returns一直是0：

请问，可能是什么原因？

[outongyiLv]

I meet this problem

this is because the author use the LlamaForCausalLM, if you do not use the llama model, you will meet this problem, i solve by replace the LlamaForCausalLM to AutoModelForCausalLM.