ingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;color: rgb(63, 63, 63);">
ingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 0.1em;color: rgb(63, 63, 63);">在这个故事中,我们将深入探讨来自 Phi-4、GPT-4o 和 o1 等主要参与者的一些最新 AI 发展。
ingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 0.1em;color: rgb(63, 63, 63);">2024年1月7日,微软在Hugging Face上发布了小型语言模型(SLM)“Phi-4”。该模型根据 MIT 许可证提供,可以自由使用和修改用于商业用途。
ingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 0.1em;color: rgb(63, 63, 63);">Phi\-4是微软“Phi”系列中的最新模型,虽然它体积小,但在某些领域表现出色。特别是在数学推理方面,它的准确性超过了大规模语言模型(LLMs),并因其在专业任务中的应用而受到关注。
ingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 0.1em;color: rgb(63, 63, 63);">Phi-4 通过高质量数据和新的后训练技术提高了准确性,并已发布到 Azure AI Foundry 和 Hugging Face。
ingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 0.1em;color: rgb(63, 63, 63);">模型大小为 14B,如果是 4 位量化模型,应该可以在不错的 GPU 上运行。Meta 的 Llama3.2 和 Google 的 Gemma2 这两个开放模型的最小模型分别为 1B 和 2B,因此这比它们大一个尺寸。
ingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 0.1em;color: rgb(63, 63, 63);">我不确定微软对“小型”的定义是什么,但无论如何,这是一个可以公开和轻松使用的东西。
ingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 0.1em;color: rgb(63, 63, 63);">在本逐步指南中,我们将介绍 Phi-4 是什么,Phi-4 的特殊功能是什么,如何在本地使用 Phi-4,以及为什么它的表现优于 LLM。 ingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;padding-left: 12px;color: rgb(63, 63, 63);">什么是 Phi-4
ingFang SC", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;font-size: 15px;letter-spacing: 0.1em;color: rgb(63, 63, 63);">Phi-4 是一种尝试展示与传统 AI 模型不同的方法。它的参数数量相对较少,仅为 140 亿,同时在性能上优于大型 AI 模型,如谷歌的 Gemini Pro 1.5\。特别引人注目的是它出色的数学推理能力。在由美国数学协会 (MAA) 主办的美国数学竞赛 (AMC) 中,Phi-4 在需要逻辑思维的问题上表现优于大型模型,例如解决复杂方程和数学证明。
这一结果显示了在科学研究、工程和金融建模等需要准确数值计算的领域中的应用潜力。
Phi-4 的特殊特点是什么? Phi-4 基准测试结果如下:
• 在 MATH 和 MGSM 测试中得分超过 80%,超越了 Google Gemini Pro 和 OpenAI 的 GPT-4o-mini。 • 在数学推理能力方面表现优秀,这在金融、工程和科学研究中非常重要。 • 编程的 HumanEval 在基准测试中表现良好。 此外,Phi-4 的架构和训练过程非常强调精确性和效率。其 140 亿参数的解码器仅模型是在总计 9.8 万亿个来自协调和合成数据集的标记上进行训练的。
在训练方法方面,微软团队使用了监督微调和直接偏好优化(DPO)方法,以确保稳固的性能并解决公平性和稳定性等问题。
微软强调,Phi-4 的训练过程考虑了负责任的人工智能发展,并经过广泛的安全评估,包括攻击测试,以降低偏见、有害内容和虚假信息的风险。
本地安装 Hugging Face 中的模型相当大,约 30GB,在我的机器上运行有点困难。
我在想是否可以通过应用 4 位量化或其他方法让它在我的机器上运行,我发现 Ollama 有一个已经量化到 Q4_K_M 的模型,因此我非常感激能够借用它,并想分享我在使用过程中记录的工作笔记。
下载 ollamarunphi4 注意: 当下载的模型扩展到内存中时,大小会略超过10GB,因此如果您在CPU上运行它,您需要至少有那么多的可用内存空间;如果您在GPU上运行它,您需要至少那么多的GPU内存,否则会出现错误。
如果模型成功下载,您应该会看到以下提示:
>>>Sendamessage(/?forhelp) 当您看到这个提示时,尝试发送一条消息。如果您使用的是GPU加速环境,您可能会很快得到响应。
>>>Hello! Hello! Please let me know how I can assist you. If you haveanyquestions, feel free to ask. >>>Send a message (/?forhelp)有了这个,我现在可以在本地使用Phi-4了!这非常有帮助。
现在,您可以使用Ollama轻松地在自己的机器上运行Phi-4!它甚至带有REST API!
这意味着具有相当于大约70B参数的LLM性能的模型可以在便携式PC上运行,而Ollama的REST API功能可以快速设置一个环境,以便可以从其他应用程序和库本地调用它。我期待在未来以各种方式利用这一点。
Phi-4 vs GPT-4o vs 和 o1 让我们比较 o1 和GPT-4o ,并验证 Phi-4 的性能。
接下来,我想提出一些应用问题,并验证结果是否与 LLMs GPT-4o 和 o1 的结果不同,以及它们是否更简单。
需要检查的应用问题如下:
1.计算以下序列的第 20 个值:1, 1, 2, 3, 5, 8, …. 2.一家银行提供 5% 的年利率,按半年复利计算。如果存入 100 万美元,存 10 年后最终金额是多少? 3. 如果一个长方体的体积是 120 立方厘米,高是 6 厘米,宽是 4 厘米,那么它的长度是多少? 问题 1 :将测试你的准确性、简洁性 和响应时间 。
问题 2 :将测试你计算利率的准确性。
问题 3 :将测试你的几何推理能力 。
接下来,让我们尝试解决这个问题,并在提示中输入以下内容:
The sequence givenisthe Fibonacci sequence, where each termisthesumof the two preceding ones, startingwith1and1.The first few terms are: ( F_1 =1) ( F_2 =1) ( F_3 = F_1 + F_2 =1+1=2) ( F_4 = F_2 + F_3 =1+2=3) ( F_5 = F_3 + F_4 =2+3=5) ( F_6 = F_4 + F_5 =3+5=8) We need to find the 20th term, ( F_{20} ). Continuing the sequence: ( F_7 = F_5 + F_6 =5+8=13) ( F_8 = F_6 + F_7 =8+13=21) ( F_9 = F_7 + F_8 =13+21=34) ( F_{10} = F_8 + F_9 =21+34=55) ( F_{11} = F_9 + F_{10} =34+55=89) ( F_{12} = F_{10} + F_{11} =55+89=144) ( F_{13} = F_{11} + F_{12} =89+144=233) ( F_{14} = F_{12} + F_{13} =144+233=377) ( F_{15} = F_{13} + F_{14} =233+377=610) ( F_{16} = F_{14} + F_{15} =377+610=987) ( F_{17} = F_{15} + F_{16} =610+987=1597) ( F_{18} = F_{16} + F_{17} =987+1597=2584) ( F_{19} = F_{17} + F_{18} =1597+2584=4181) ( F_{20} = F_{18} + F_{19} =2584+4181=6765) Thus, the 20th value of the sequenceis( \boxed{6765} ).Phi-4 给定的序列是斐波那契数列,其中每个项是前两个项的和,从1和1开始。前几个项是: ( F_1 =1) ( F_2 =1) ( F_3 = F_1 + F_2 =1+1=2) ( F_4 = F_2 + F_3 =1+2=3) ( F_5 = F_3 + F_4 =2+3=5) ( F_6 = F_4 + F_5 =3+5=8) 我们需要找到第20项,( F_{20} )。 继续序列: ( F_7 = F_5 + F_6 =5+8=13) ( F_8 = F_6 + F_7 =8+13=21) ( F_9 = F_7 + F_8 =13+21=34) ( F_{10} = F_8 + F_9 =21+34=55) ( F_{11} = F_9 + F_{10} =34+55=89) ( F_{12} = F_{10} + F_{11} =55+89=144) ( F_{13} = F_{11} + F_{12} =89+144=233) ( F_{14} = F_{12} + F_{13} =144+233=377) ( F_{15} = F_{13} + F_{14} =233+377=610) ( F_{16} = F_{14} + F_{15} =377+610=987) ( F_{17} = F_{15} + F_{16} =610+987=1597) ( F_{18} = F_{16} + F_{17} =987+1597=2584) ( F_{19} = F_{17} + F_{18} =1597+2584=4181) ( F_{20} = F_{18} + F_{19} =2584+4181=6765) 因此,序列的第20个值是 ( \boxed{6765} )。GPT-4o 给定的序列是**斐波那契数列**,其中每个项是前两项的和: \[F(n) = F(n-1) + F(n-2) \quad \text{当 } n \geq3\] 且 \( F(1) =1\) 和 \( F(2) =1\)。 我们需要计算这个序列的**第20个值**。 前几项是: 1.\( F(1) =1\) 2.\( F(2) =1\) 3.\( F(3) =2\) 4.\( F(4) =3\) 5.\( F(5) =5\) 6.\( F(6) =8\) 7.\( F(7) =13\) 8.\( F(8) =21\) 9.\( F(9) =34\) 10.\( F(10) =55\) 11.\( F(11) =89\) 12.\( F(12) =144\) 13.\( F(13) =233\) 14.\( F(14) =377\) 15.\( F(15) =610\) 16.\( F(16) =987\) 17.\( F(17) =1597\) 18.\( F(18) =2584\) 19.\( F(19) =4181\) 20.\( F(20) =6765\) 因此,序列中的第20个值是:6765o1 The given sequenceisthe Fibonacci sequence, defined by: \[F_1 =1, \quad F_2 =1, \quad \text{and} \quad F_{n} = F_{n-1} + F_{n-2} \text{for} n >2.\] To find the 20th term, let's list them out (or recall the known values): 1. \( F_1 = 1 \) 2. \( F_2 = 1 \) 3. \( F_3 = 2 \) 4. \( F_4 = 3 \) 5. \( F_5 = 5 \) 6. \( F_6 = 8 \) 7. \( F_7 = 13 \) 8. \( F_8 = 21 \) 9. \( F_9 = 34 \) 10. \( F_{10} = 55 \) 11. \( F_{11} = 89 \) 12. \( F_{12} = 144 \) 13. \( F_{13} = 233 \) 14. \( F_{14} = 377 \) 15. \( F_{15} = 610 \) 16. \( F_{16} = 987 \) 17. \( F_{17} = 1597 \) 18. \( F_{18} = 2584 \) 19. \( F_{19} = 4181 \) 20. \( F_{20} = 6765 \) Therefore, the 20th value in the sequence is: 6765所有模型都能给出正确答案,但Phi-4的解法适合没有强数学背景的人。得到答案所需的时间差别不大。
让我们尝试解决第二个问题,并在您的聊天机器人中使用这个提示
A bank offers an annual interest rate of5%, compounded semi-annually. Whatisthe final amountif1million dollarisdepositedfor10years?Phi-4 To calculate the final amount when interestiscompounded semi-annually, we can use the compound interest formula: [ A = P \left(1+ \frac{r}{n}\right)^{nt} ] where: ( A )isthe final amount ( P )isthe principal amount (initial deposit) ( r )isthe annual interest rate (indecimal form) ( n )isthe number of times the interestiscompounded per year ( t )isthe number of years For this problem: ( P =1,000,000) dollars ( r =0.05) (5%asa decimal) ( n =2) (compounded semi-annually) ( t =10) years Plugging the values into the formula: [ A =1,000,000\left(1+ \frac{0.05}{2}\right)^{2\times10} ] [ A =1,000,000\left(1+0.025\right)^{20} ] [ A =1,000,000\times (1.025)^{20} ] Now, calculate ( (1.025)^{20} ): [ (1.025)^{20} \approx1.638616] Therefore: [ A \approx1,000,000\times1.638616] [ A \approx1,638,616] The final amount after10yearsisapproximately $1,638,616.GPT-4o 要计算复利的最终金额,可以使用以下公式: \[ A = P \left(1+ \frac{r}{n}\right)^{nt} \] 其中: - \( A \) 是最终金额, - \( P \) 是本金(初始存款), - \( r \) 是年名义利率(以小数表示), - \( n \) 是每年复利的次数, - \( t \) 是年数。 给定: - \( P =1,000,000\) 美元, - \( r =0.05\)(5%), - \( n =2\)(因为利息是半年复利), - \( t =10\)。 将这些值代入公式: \[ A =1,000,000\left(1+ \frac{0.05}{2}\right)^{2\times10} \] \[ A =1,000,000\left(1+0.025\right)^{20} \] \[ A =1,000,000\times (1.025)^{20} \] 现在计算 \( (1.025)^{20} \): \[ (1.025)^{20} \approx1.6386164403\] 现在找到最终金额 \( A \): \[ A \approx1,000,000\times1.6386164403\] \[ A \approx1,638,616.44\] 10年后的最终金额将约为 \$1,638,616.44。o1 要计算一个 \$1,000,000存款在年利率为5%、每年复利两次的情况下,经过10年后的最终金额,可以使用复利公式: \[ A = P \left(1+ \frac{r}{n}\right)^{nt} \] 其中: - \( A \) = n 年后累积的金额,包括利息。 - \( P \) = 本金 (\$1,000,000)。 - \( r \) = 年利率(小数)(5% 或0.05)。 - \( n \) = 每年复利的次数(每年两次为2)。 - \( t \) = 投资的年数(10年)。 代入数值: \[ A =1,000,000\times \left(1+ \frac{0.05}{2}\right)^{2\times10} =1,000,000\times (1.025)^{20} \] 计算 \( (1.025)^{20} \): \[ (1.025)^{20} \approx1.63862 \] 现在,将其乘以本金: \[ A \approx1,000,000\times1.63862=1,638,620 \] **最终金额:**经过10年,投资将增长到大约 **\$1,638,620**。• 所有模型都能够获得正确的结果。O1 的速度显著快于其他模型,仅用 11.38 毫秒完成任务。GPT-4o 和Phi-4 的速度较慢,其中 Phi-4 的时间稍长于 GPT-4o。 推理 最后,我们将测试几何推理能力的测试。请在您的聊天机器人中使用此提示
If the volume of a rectangular prismis120cubic centimetres, its heightis6cm,andits widthis4cm, whatisits length?Phi-4 要找到矩形棱柱的长度,我们可以使用矩形棱柱体积的公式: [ \text{Volume} = \text{Length} \times \text{Width} \times \text{Height} ] 已知: 体积 =120立方厘米 高度 =6cm 宽度 =4cm 将已知值代入公式: [120= \text{Length} \times4\times6] 简化方程: [120= \text{Length} \times24] 要解出长度,将两边都除以24: [ \text{Length} = \frac{120}{24} ] 计算除法: [ \text{Length} =5] 因此,矩形棱柱的长度为5cm。GPT-4o To find the length of the rectangular prism, you can use the formulaforthe volume of a rectangular prism: \[ \text{Volume} = \text{length} \times \text{width} \times \text{height} \] You are given the volume (120cubic centimeters), the height (6cm),andthe width (4cm). Plug these values into the formula: \[120= \text{length} \times4\times6\] First, calculate \(4\times6\): \[120= \text{length} \times24\] Next, solveforthe length by dividing both sides of the equation by24: \[ \text{length} = \frac{120}{24} \] \[ \text{length} =5\] Thus, the length of the rectangular prismis5centimeters.o1 要找到矩形棱柱的**长度**,可以使用矩形棱柱体积的公式: \[ \text{体积} = \text{长度} \times \text{宽度} \times \text{高度} \] 给定: - **体积 (V)** =120立方厘米 (cm³) - **高度 (h)** =6厘米 (cm) - **宽度 (w)** =4厘米 (cm) **步骤1:写下包含已知值的公式。** \[ 120= \text{长度} \times4\times6 \] **步骤2:简化方程以求解长度。** \[ \text{长度} = \frac{120}{4\times6} \] \[ \text{长度} = \frac{120}{24} \] \[ \text{长度} =5\text{ cm} \] **答案:** 矩形棱柱的长度是**5厘米**。所有模型都给出了正确的结果,但O1和Phi-4在推理方面更优秀。它们逐步解决问题,就像人类思考一样。我很惊讶GPT-4o在推理方面不如O1和Phi-4。
结论: 我比较了Phi-4与GPT-4o和O1,发现速度几乎相同。Phi-4是开源的,我更喜欢开源而不是GPT-4o和O1。尽管O1速度更快、性能更好,但我还是选择了Phi-4。我认为Phi-4与其他模型之间没有太大差别,而且Phi-4足够小,可以在我的笔记本电脑上运行,这一点非常重要,因为我没有一台强大的计算机。