Socraticlm

🚀 SocraticLM模型

SocraticLM是一個基於Qwen2.5-Math-7B-Instruct微調的模型，專為教育場景設計。它能以蘇格拉底式引導幫助學生解決數學問題，也可直接解答數學問題，支持中英文。

🚀 快速開始

模型簡介

本模型是在Qwen2.5-Math-7B-Instruct基礎上，使用SocraTeach數據集進行微調得到的，是SocraticLM的具體實現。

預期用途與限制

SocraticLM主要用於教育場景，當學生在學習解決數學問題遇到困難時，它能提供蘇格拉底式的引導。同時，該模型也可以直接解決數學問題。此模型主要支持英文和中文。

安裝指南

此部分文檔未提及具體安裝步驟，因此跳過該章節。

💻 使用示例

基礎用法

使用Huggingface transformers庫

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("CogBase-USTC/SocraticLM")
model = AutoModelForCausalLM.from_pretrained(
    "CogBase-USTC/SocraticLM",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

### 數學問題求解 ###
messages = [
    {"role": "system", "content" : "Please analyse and solve the following problem step by step."},
    {"role": "user", "content": "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?"},
]

### 蘇格拉底式引導 ###
# messages = [
#     {"role": "system", "content" : "You are a Socratic teacher, please guide me to solve the [Problem] with heuristic questions based on the following information. \n"},
#     {"role": "user", "content": "[Problem] Debelyn, Christel, and Andrena collect dolls. Debelyn had 20 dolls before she gave Andrena 2 dolls. Christel had 24 dolls before giving Andrena 5 dolls. After all the gifts, Andrena now has 2 more dolls  than Christel, how many more dolls does andrena have now than Debelyn? [Answer] 3 [Analysis] Debelyn had 20 - 2 = 18 dolls left after giving out 2 dolls to Christel. Christel had 24 + 2 = 26 dolls after receiving 2 dolls from Debelyn. Christel had 24 - 5 = 19 dolls after giving Andrena 5 dolls. So, Andrena has 19 +2 = 21 dolls now. Therefore, Andrena has 21 - 18 = 3 more dolls than Debelyn."},
# ]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer.encode(prompt, return_tensors="pt")
outputs = model.generate(input_ids=inputs.to(model.device), max_new_tokens=4096)
print(tokenizer.decode(outputs[0]))

使用vLLM庫

from vllm import LLM, SamplingParams

llm = LLM(model=r'CogBase-USTC/SocraticLM',
          tokenizer=r'CogBase-USTC/SocraticLM',
          trust_remote_code=True,
          tensor_parallel_size=1,
          gpu_memory_utilization=0.99,
          enable_chunked_prefill=True,
          max_num_batched_tokens=512,
          max_num_seqs=128)
sampling_params = SamplingParams(temperature=0, max_tokens=4096, seed=42)


def print_outputs(outputs):
    for output in outputs:
        prompt = output.prompt
        generated_text = output.outputs[0].text
        print(f"Generated text: {generated_text!r}")
    print("-" * 80)


print("=" * 80)

### 數學問題求解 ###
conversation = [
    {
        "role": "system",
        "content": "Please analyse and solve the following problem step by step."
    },
    {
        "role": "user", 
        "content": "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?"
    },
]

### 蘇格拉底式引導 ###
# conversation = [
#     {
#         "role": "system",
#         "content": "You are a Socratic teacher, please guide me to solve the [Problem] with heuristic questions based on the following information. \n"
#     },
#     {
#         "role": "user", 
#         "content": "[Problem] Debelyn, Christel, and Andrena collect dolls. Debelyn had 20 dolls before she gave Andrena 2 dolls. Christel had 24 dolls before giving Andrena 5 dolls. After all the gifts, Andrena now has 2 more dolls  than Christel, how many more dolls does andrena have now than Debelyn? [Answer] 3 [Analysis] Debelyn had 20 - 2 = 18 dolls left after giving out 2 dolls to Christel. Christel had 24 + 2 = 26 dolls after receiving 2 dolls from Debelyn. Christel had 24 - 5 = 19 dolls after giving Andrena 5 dolls. So, Andrena has 19 +2 = 21 dolls now. Therefore, Andrena has 21 - 18 = 3 more dolls than Debelyn."
#     },
# ]

outputs = llm.chat(conversation,
                   sampling_params=sampling_params,
                   use_tqdm=False,
                )
print_outputs(outputs)

詳細文檔

訓練超參數

訓練過程中使用的超參數如下：

• 學習率（learning_rate）：1e-05
• 訓練批次大小（train_batch_size）：8
• 隨機種子（seed）：42
• 分佈式類型（distributed_type）：多GPU
• 設備數量（num_devices）：4
• 總訓練批次大小（total_train_batch_size）：32
• 總評估批次大小（total_eval_batch_size）：32
• 優化器（optimizer）：使用OptimizerNames.ADAMW_TORCH，其中beta1=0.9，beta2=0.999，epsilon=1e-08，無額外優化器參數
• 學習率調度器類型（lr_scheduler_type）：餘弦
• 學習率調度器熱身步數（lr_scheduler_warmup_steps）：20

框架版本

• Transformers：4.46.2
• Pytorch：2.5.1+cu124
• Datasets：3.1.0
• Tokenizers：0.20.3

📄 許可證

許可證類型：其他（other）