|
--- |
|
license: apache-2.0 |
|
datasets: |
|
- KingNish/reasoning-base-20k |
|
language: |
|
- en |
|
- zh |
|
base_model: |
|
- Qwen/Qwen2.5-1.5B |
|
--- |
|
|
|
## Uses |
|
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
model_name = "/root/app/Reason/checkpoints" |
|
|
|
model = AutoModelForCausalLM.from_pretrained( |
|
model_name, |
|
torch_dtype="auto", |
|
device_map="auto" |
|
) |
|
tokenizer = AutoTokenizer.from_pretrained(model_name) |
|
|
|
from typing import List, Dict |
|
def new_apply_chat_template(history:List[Dict[str, str]], add_reasoning_generation_prompt:bool=True, add_assistant_generation_prompt:bool=False): |
|
if add_reasoning_generation_prompt: |
|
return "".join([f"<|im_start|>{i['role']}\n{i['content']}<|im_end|>\n" for i in history]) + "<|im_start|><|reasoning|>\n" |
|
if add_assistant_generation_prompt: |
|
return "".join([f"<|im_start|>{i['role']}\n{i['content']}<|im_end|>\n" for i in history]) + "<|im_start|>assistant\n" |
|
|
|
|
|
from IPython.display import Markdown, display |
|
device = "cuda" |
|
history = [] |
|
history.append({"role": "system", "content": "You are a helpful assistant"}) |
|
while True: |
|
question = input('User：' + '\n') |
|
print(question) |
|
print('\n') |
|
history.append({"role": "user", "content": question}) |
|
|
|
input_text = new_apply_chat_template( |
|
history, |
|
add_reasoning_generation_prompt=True |
|
) |
|
model_inputs = tokenizer([input_text], return_tensors="pt").to(device) |
|
|
|
if model_inputs.input_ids.size()[1]>32000: |
|
break |
|
|
|
generated_ids = model.generate( |
|
model_inputs.input_ids, |
|
max_new_tokens=3000 |
|
) |
|
|
|
if len(generated_ids)>32000: |
|
break |
|
|
|
generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)] |
|
|
|
reasoning_response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] |
|
history.append({"role": "<|reasoning|>", "content": reasoning_response}) |
|
print('reasoning:\n') |
|
#print(response) |
|
display(Markdown(reasoning_response)) |
|
print("------------") |
|
print('\n') |
|
|
|
input_text = new_apply_chat_template( |
|
history, |
|
add_assistant_generation_prompt=True |
|
) |
|
model_inputs = tokenizer([input_text], return_tensors="pt").to(device) |
|
|
|
if model_inputs.input_ids.size()[1]>32000: |
|
break |
|
|
|
generated_ids = model.generate( |
|
model_inputs.input_ids, |
|
max_new_tokens=3000 |
|
) |
|
|
|
if len(generated_ids)>32000: |
|
break |
|
|
|
generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)] |
|
|
|
assistant_response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0] |
|
history.append({"role": "assistant", "content": assistant_response}) |
|
print('assistant:\n') |
|
display(Markdown(assistant_response)) |
|
print("------------") |
|
|
|
print("超过模型字数上线，已退出") |
|
|
|
|
|
|
|
在模型中加入 `<|reasoning|>` 特殊标记： |
|
|
|
1. 训练数据格式修改 |
|
- 在 `prepare_conversation_template` 方法中，对话模板增加了 reasoning 部分 |
|
- 对话流程变成: system → user → reasoning → assistant |
|
- 这让模型能够学习到推理过程和最终答案之间的关系 |
|
|
|
2. 模型架构变化 |
|
- 通过 `add_special_tokens` 方法添加了新的特殊标记 `<|reasoning|>` |
|
- 扩展了模型的词表大小(通过 `resize_token_embeddings`) |
|
- 让模型能够识别和生成这个特殊标记 |
|
|
|
3. 训练目标的改变 |
|
- 模型不仅要学习生成答案,还要学习生成推理过程 |
|
- 在 `create_answer_mask` 中，模型需要同时预测 reasoning 和 assistant 部分 |
|
|
|
4. 推理能力增强 |
|
- 模型被训练去显式地展示推理过程 |
|
- 有助于生成更可解释的答案 |
|
- 可以帮助模型形成更结构化的思维链路 |
|
|
|
让模型在回答问题时不仅给出答案，还能展示推理过程，提高模型回答的可解释性和可靠性。 |
|
|
|
|
|
与不改变词表的本质区别在于： |
|
|
|
1. 不改词表的方式： |
|
- `<|reasoning|>` 会被拆分成普通token序列来处理 |
|
- 模型将其视为普通文本，类似于看到 "reasoning" 这个词 |
|
- 占用更多token位置(因为被拆分) |
|
- 没有特殊的语义表示学习 |
|
|
|
2. 添加到词表的方式： |
|
- `<|reasoning|>` 作为单个特殊token处理 |
|
- 模型可以学习到这个标记专门的语义表示 |
|
- 只占用一个token位置 |
|
- 能够建立独特的语义空间 |
|
|
|
具体影响： |
|
``` |
|
# 假设tokenizer编码示例 |
|
# 不加入词表: |
|
"<|reasoning|>" -> [1234, 5678, 9012] # 被拆分成多个基础token |
|
|
|
# 加入词表: |
|
"<|reasoning|>" -> [32000] # 单个特殊token |
|
``` |
|
|
|
从训练效果看： |
|
- 加入词表能让模型更好地识别和区分推理部分的边界 |
|
- 不加入词表依然可以起到标记作用，但模型需要从上下文学习这种模式 |
|
- 加入词表的方式通常效果更好，因为给了模型更明确的语义信号 |
|
|
|
|
