Release Mtrini-27B-Tellus GGUF models

.gitattributes

@@ -34,3 +34,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
+Mtrini-27B-Tellus-F16-FINAL.gguf filter=lfs diff=lfs merge=lfs -text
+Mtrini-27B-Tellus-IQ2_XS.gguf filter=lfs diff=lfs merge=lfs -text
+Mtrini-27B-Tellus-IQ2_XS.imatrix filter=lfs diff=lfs merge=lfs -text

Mtrini-27B-Tellus-F16-FINAL.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc07b0c18d1415d2328008fcb88e4946991e93393e2a6c74aa0ec781df91dd9b
+size 53808282112

Mtrini-27B-Tellus-IQ2_XS.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98e256c55c90f1f89ea846be1881b3e60be517667017ec2b9382a62eb0adc736
+size 9090591520

Mtrini-27B-Tellus-IQ2_XS.imatrix

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0676b166f743b3c49085f969c0cb8d74bd3d7fdfca5b1fb43217570de8c91a13
+size 13642688

README.md

@@ -1,62 +1,175 @@
 ---
-base_model: Qwen/Qwen3.8-27B
-library_name: peft
-model_name: mtrini-27b-tellus-adapter
-tags:
-- base_model:adapter:Qwen/Qwen3.8-27B
-- lora
-- sft
-- transformers
-- trl
-licence: license
+library_name: gguf
+license: apache-2.0
 pipeline_tag: text-generation
+tags:
+- gguf
+- qwen
+- qwen3
+- qwen3.5
+- coding
+- mathematics
+- reasoning
+- darija
+- moroccan-arabic
+- arabic
+- multilingual
+- compiwerai
+- mtrini
+- iq2_xs
 ---
 
-# Model Card for mtrini-27b-tellus-adapter
+# Mtrini-27B-Tellus
+
+**Mtrini-27B-Tellus** is an open model developed by **CompiwerAI**.
+
+Mtrini is designed for coding, mathematical reasoning, general reasoning, Arabic, Moroccan Arabic and Moroccan Darija.
+
+## Model
+
+- **Model:** Mtrini-27B-Tellus
+- **Organization:** CompiwerAI
+- **Base:** Qwen3.8-27B
+- **Parameters:** approximately 27B
+- **Architecture:** Qwen3.5-compatible GGUF metadata
+- **Training context:** 4096 tokens
+- **Training steps:** 700
+- **Effective batch size:** 16
+- **LoRA rank:** 32
+- **LoRA alpha:** 64
+- **LoRA dropout:** 0.05
+- **Learning rate:** 0.00015
+- **Compute dtype:** BF16
+- **QLoRA quantization:** 4-bit NF4
+
+## Training data
+
+| Dataset | Share | Examples |
+|---|---:|---:|
+| OpenCodeInstruct | 35% | 3,920 |
+| Magicoder | 10% | 1,120 |
+| OpenR1-Math | 25% | 2,800 |
+| Moroccan Darija | 20% | 2,240 |
+| Aya Arabic / Moroccan Arabic | 10% | 1,120 |
+| **Total** | **100%** | **11,200** |
+
+## Training result
+
+- Global step: 700
+- Final reported loss: 0.4302458722250802
+- Runtime: approximately 3.15 hours
+- Epoch: 1
+- GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition
+
+## Quantized model
+
+### Mtrini-27B-Tellus-IQ2_XS.gguf
+
+Recommended for practical local inference.
+
+- Size: approximately 8.47 GiB
+- Quantization: IQ2_XS
+- Bits per weight: 2.3125 bpw
+- Tensor count: 851
+- GGUF architecture: qwen35
+- Block count: 64
+
+The IQ2_XS quantization was generated using an importance matrix calculated from a dedicated calibration corpus.
+
+## F16 model
+
+### Mtrini-27B-Tellus-F16-FINAL.gguf
+
+- Size: approximately 50.11 GiB
+- Format: F16 GGUF
 
-This model is a fine-tuned version of [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B).
-It has been trained using [TRL](https://github.com/huggingface/trl).
+This version is intended for systems with sufficient memory.
 
-## Quick start
+## Importance matrix
 
-```python
-from transformers import pipeline
+File:
 
-question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
-generator = pipeline("text-generation", model="None", device_map="auto")
-output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
-print(output["generated_text"])
+`Mtrini-27B-Tellus-IQ2_XS.imatrix`
+
+Generation settings:
+
+- Calibration context: 512
+- Threads: 20
+- Calibration chunks: 32
+- Calibration PPL: approximately 2.4629
+
+## Calibration data
+
+File:
+
+`mtrini-calibration-large.txt`
+
+This corpus was used during importance-matrix generation.
+
+## GGUF metadata
+
+- GGUF magic: valid
+- Tensor count: 851
+- Architecture: qwen35
+- Block count: 64
+- Embedding size: 5120
+- FFN size: 17408
+- Attention heads: 24
+- KV heads: 4
+
+## llama.cpp
+
+Example:
+
+```bash
+llama-cli \
+  -m Mtrini-27B-Tellus-IQ2_XS.gguf \
+  -ngl 99 \
+  -c 4096 \
+  -p "What is 27 multiplied by 8?"
 ```
 
-## Training procedure
+## Basic verification
+
+Prompt:
+
+> What is 27 multiplied by 8? Answer with only the number.
+
+Observed CPU test output:
+
+**216**
+
+## Intended uses
+
+- Code generation
+- Code debugging
+- Mathematical reasoning
+- General reasoning
+- Moroccan Darija
+- Arabic
+- Multilingual text
+- Local inference
+- GGUF-compatible inference
 
- 
+## Limitations
 
+Mtrini-27B-Tellus can generate incorrect information, incorrect code, hallucinations, or unreliable reasoning.
 
+Generated code and important factual outputs should be validated.
 
-This model was trained with SFT.
+The IQ2_XS version makes a substantial size/quality tradeoff in exchange for a much smaller model footprint.
 
-### Framework versions
+## Files
 
-- PEFT 0.21.0
-- TRL: 1.14.0
-- Transformers: 5.17.0
-- Pytorch: 2.11.0
-- Datasets: 5.0.1
-- Tokenizers: 0.23.2
+- `Mtrini-27B-Tellus-IQ2_XS.gguf`
+- `Mtrini-27B-Tellus-IQ2_XS.imatrix`
+- `Mtrini-27B-Tellus-F16-FINAL.gguf`
+- `mtrini-calibration-large.txt`
+- `README.md`
 
-## Citations
+## Credits
 
+**CompiwerAI**
 
+**Building AI For Everyone.**
 
-Cite TRL as:
-    
-```bibtex
-@software{vonwerra2020trl,
-  title   = {{TRL: Transformers Reinforcement Learning}},
-  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
-  license = {Apache-2.0},
-  url     = {https://github.com/huggingface/trl},
-  year    = {2020}
-}
-```

mtrini-calibration-large.txt

@@ -0,0 +1,2070 @@
+[Calibration sample 0-0]
+Write a Python function that determines whether a string is a palindrome.
+The function should ignore spaces, punctuation, and capitalization.
+Explain the algorithm and its time complexity.
+
+
+[Calibration sample 0-1]
+Implement binary search in Python for a sorted list of integers.
+Return the index of the target or -1 when it is not present.
+Explain why binary search runs in logarithmic time.
+
+
+[Calibration sample 0-2]
+Write a Python function that counts the frequency of every word
+in a text. Normalize capitalization and ignore punctuation.
+Return the words ordered from most frequent to least frequent.
+
+
+[Calibration sample 0-3]
+Given a list of integers, find the maximum subarray sum using
+Kadane's algorithm. Explain why the algorithm works and give its
+time and space complexity.
+
+
+[Calibration sample 0-4]
+Implement a stack using two queues. Explain each operation and
+the tradeoff between the possible implementations.
+
+
+[Calibration sample 0-5]
+Write a function that checks whether two strings are anagrams.
+The solution should run in O(n) time using a frequency dictionary.
+
+
+[Calibration sample 0-6]
+Explain the difference between a process and a thread in an
+operating system. Give an example where multithreading is useful.
+
+
+[Calibration sample 0-7]
+Write a recursive function to calculate the factorial of an integer.
+Handle zero correctly and explain the base case.
+
+
+[Calibration sample 0-8]
+Explain what an API is and show a small Python example of sending
+a JSON request to an HTTP endpoint.
+
+
+[Calibration sample 0-9]
+Design a simple REST API for a chat application. Describe endpoints
+for creating conversations, sending messages, and retrieving history.
+
+
+[Calibration sample 0-10]
+Solve the equation 3x + 7 = 31. Show every algebraic step.
+
+
+[Calibration sample 0-11]
+A rectangle has a length of 12 centimeters and a width of 7 centimeters.
+Calculate its area and perimeter.
+
+
+[Calibration sample 0-12]
+If a train travels 240 kilometers in 3 hours, calculate its average
+speed in kilometers per hour.
+
+
+[Calibration sample 0-13]
+A number is multiplied by 4 and then increased by 9 to produce 37.
+Find the original number.
+
+
+[Calibration sample 0-14]
+There are 5 red balls and 3 blue balls in a bag. What is the probability
+of randomly selecting a blue ball? Explain your reasoning.
+
+
+[Calibration sample 0-15]
+A shop gives a 20 percent discount on a product priced at 250 dirhams.
+Calculate the discount and the final price.
+
+
+[Calibration sample 0-16]
+If x + y = 18 and x - y = 6, solve for x and y using a system of equations.
+
+
+[Calibration sample 0-17]
+Explain why the square root of 2 is irrational using a proof by contradiction.
+
+
+[Calibration sample 0-18]
+A sequence begins 2, 4, 8, 16, 32. Describe the pattern and calculate
+the next three terms.
+
+
+[Calibration sample 0-19]
+A triangle has a base of 10 centimeters and a height of 8 centimeters.
+Calculate its area and explain the formula.
+
+
+[Calibration sample 0-20]
+A farmer has chickens and cows. There are 20 animals and 56 legs.
+Determine how many chickens and cows there are. Explain the reasoning.
+
+
+[Calibration sample 0-21]
+Three boxes are labeled apples, oranges, and mixed. Every label is wrong.
+You may take one fruit from one box. Explain how to correctly relabel
+all three boxes.
+
+
+[Calibration sample 0-22]
+A clock shows 3:15. What is the smaller angle between the hour hand
+and minute hand? Explain carefully.
+
+
+[Calibration sample 0-23]
+You have eight identical-looking balls and one is heavier.
+Using a balance scale only twice, explain how to find the heavier ball.
+
+
+[Calibration sample 0-24]
+A sequence contains 1, 1, 2, 3, 5, 8, 13. Identify the rule and
+calculate the next five values.
+
+
+[Calibration sample 0-25]
+Explain artificial intelligence to a ten-year-old using a simple
+real-world analogy. Avoid technical jargon where possible.
+
+
+[Calibration sample 0-26]
+Explain the difference between machine learning and traditional
+programming using a clear example.
+
+
+[Calibration sample 0-27]
+What is a neural network? Explain neurons, weights, training,
+and inference in simple but technically accurate language.
+
+
+[Calibration sample 0-28]
+Explain why training data quality can matter as much as training
+data quantity when building a language model.
+
+
+[Calibration sample 0-29]
+Describe the difference between an open-source model and an
+open-weight model. Explain why the distinction matters.
+
+
+[Calibration sample 0-30]
+اشرح مفهوم الذكاء الاصطناعي بطريقة بسيطة لطفل صغير، مع إعطاء أمثلة
+من الحياة اليومية وكيف يمكن للحاسوب أن يتعلم من البيانات.
+
+
+[Calibration sample 0-31]
+اشرح الفرق بين البرمجة التقليدية وتعلم الآلة، وقدم مثالا بسيطا
+يوضح كيف يتعلم النموذج من البيانات.
+
+
+[Calibration sample 0-32]
+حل المع��دلة 3س + 7 = 31 واشرح جميع الخطوات بطريقة واضحة.
+
+
+[Calibration sample 0-33]
+ما هو النموذج اللغوي الكبير؟ اشرح كيف يتم تدريبه وكيف يستطيع
+توليد النصوص بعد التدريب.
+
+
+[Calibration sample 0-34]
+اشرح أهمية البيانات الجيدة عند تدريب نماذج الذكاء الاصطناعي
+ولماذا لا تعني كمية البيانات الكبيرة دائما جودة أفضل.
+
+
+[Calibration sample 0-35]
+شرح ليا شنو هو الذكاء الاصطناعي بطريقة بسيطة وبأمثلة من الحياة
+اليومية ديالنا فالمغرب.
+
+
+[Calibration sample 0-36]
+شنو الفرق بين البرمجة العادية وتعلم الآلة؟ شرحها ليا بطريقة سهلة
+وبمثال واضح.
+
+
+[Calibration sample 0-37]
+إلى كان عندنا 3x + 7 = 31، شحال كتساوي x؟ وشرح ليا المراحل وحدة بوحدة.
+
+
+[Calibration sample 0-38]
+شرح ليا شنو هو الموديل ديال اللغة الكبير وكيفاش كيتعلم من الداتا.
+
+
+[Calibration sample 0-39]
+علاش الداتا النظيفة والمزيانة مهمة فالتدريب ديال موديلات الذكاء
+الاصطناعي؟ وعلاش ماشي ديما الكمية الكبيرة هي الأفضل؟
+
+
+[Calibration sample 0-40]
+بغيت نبني تطبيق صغير بالبايثون كيجاوب على أسئلة المستخدمين.
+شنو هي المكونات اللي خاصني وكيفاش نقدر ننظم المشروع؟
+
+
+[Calibration sample 0-41]
+شرح ليا شنو هو API وكيفاش نقدر نستعملو فمشروع ديالي بالبايثون.
+
+
+[Calibration sample 0-42]
+عندي لائحة ديال الأرقام وبغيت نلقى أكبر مجموع ديال عناصر متتابعة
+فيها. شرح ليا كيفاش نخدم بخوارزمية Kadane.
+
+
+[Calibration sample 0-43]
+شنو الفرق بين CPU و GPU؟ وعلاش الـGPU مهم بزاف فالتدريب ديال
+موديلات الذكاء الاصطناعي؟
+
+
+[Calibration sample 0-44]
+شرح ليا علاش خاصنا نديرو validation للبيانات قبل ما ندربو موديل
+ديال machine learning.
+
+
+[Calibration sample 0-45]
+Large language models are neural networks trained to predict tokens.
+During training, the model processes sequences of text and adjusts its
+parameters so that its predictions become closer to the expected next
+tokens. Modern transformer architectures use attention mechanisms to
+connect information across different positions in a sequence. During
+inference, the trained parameters remain fixed while the model generates
+new tokens step by step.
+
+
+[Calibration sample 0-46]
+Quantization reduces the numerical precision used to represent model
+weights. A model stored in BF16 may require substantially more memory
+than the same model represented with four-bit or two-bit quantization.
+Lower-bit quantization can make large models easier to deploy, although
+aggressive quantization can reduce accuracy and may require techniques
+such as importance matrices to preserve the most important information.
+
+
+[Calibration sample 0-47]
+Software engineering requires careful handling of errors, interfaces,
+dependencies, testing, and maintainability. A program that works for one
+input is not necessarily robust. Good implementations consider edge
+cases, validate assumptions, use clear abstractions, and provide useful
+error messages when something goes wrong.
+
+
+[Calibration sample 0-48]
+When evaluating an AI model, it is useful to measure different
+capabilities separately. Coding benchmarks can test program synthesis,
+mathematical datasets can test symbolic and numerical reasoning, and
+language evaluations can measure instruction following and multilingual
+performance. A single benchmark score does not completely describe the
+capabilities of a model.
+
+
+[Calibration sample 1-0]
+Write a Python function that determines whether a string is a palindrome.
+The function should ignore spaces, punctuation, and capitalization.
+Explain the algorithm and its time complexity.
+
+
+[Calibration sample 1-1]
+Implement binary search in Python for a sorted list of integers.
+Return the index of the target or -1 when it is not present.
+Explain why binary search runs in logarithmic time.
+
+
+[Calibration sample 1-2]
+Write a Python function that counts the frequency of every word
+in a text. Normalize capitalization and ignore punctuation.
+Return the words ordered from most frequent to least frequent.
+
+
+[Calibration sample 1-3]
+Given a list of integers, find the maximum subarray sum using
+Kadane's algorithm. Explain why the algorithm works and give its
+time and space complexity.
+
+
+[Calibration sample 1-4]
+Implement a stack using two queues. Explain each operation and
+the tradeoff between the possible implementations.
+
+
+[Calibration sample 1-5]
+Write a function that checks whether two strings are anagrams.
+The solution should run in O(n) time using a frequency dictionary.
+
+
+[Calibration sample 1-6]
+Explain the difference between a process and a thread in an
+operating system. Give an example where multithreading is useful.
+
+
+[Calibration sample 1-7]
+Write a recursive function to calculate the factorial of an integer.
+Handle zero correctly and explain the base case.
+
+
+[Calibration sample 1-8]
+Explain what an API is and show a small Python example of sending
+a JSON request to an HTTP endpoint.
+
+
+[Calibration sample 1-9]
+Design a simple REST API for a chat application. Describe endpoints
+for creating conversations, sending messages, and retrieving history.
+
+
+[Calibration sample 1-10]
+Solve the equation 3x + 7 = 31. Show every algebraic step.
+
+
+[Calibration sample 1-11]
+A rectangle has a length of 12 centimeters and a width of 7 centimeters.
+Calculate its area and perimeter.
+
+
+[Calibration sample 1-12]
+If a train travels 240 kilometers in 3 hours, calculate its average
+speed in kilometers per hour.
+
+
+[Calibration sample 1-13]
+A number is multiplied by 4 and then increased by 9 to produce 37.
+Find the original number.
+
+
+[Calibration sample 1-14]
+There are 5 red balls and 3 blue balls in a bag. What is the probability
+of randomly selecting a blue ball? Explain your reasoning.
+
+
+[Calibration sample 1-15]
+A shop gives a 20 percent discount on a product priced at 250 dirhams.
+Calculate the discount and the final price.
+
+
+[Calibration sample 1-16]
+If x + y = 18 and x - y = 6, solve for x and y using a system of equations.
+
+
+[Calibration sample 1-17]
+Explain why the square root of 2 is irrational using a proof by contradiction.
+
+
+[Calibration sample 1-18]
+A sequence begins 2, 4, 8, 16, 32. Describe the pattern and calculate
+the next three terms.
+
+
+[Calibration sample 1-19]
+A triangle has a base of 10 centimeters and a height of 8 centimeters.
+Calculate its area and explain the formula.
+
+
+[Calibration sample 1-20]
+A farmer has chickens and cows. There are 20 animals and 56 legs.
+Determine how many chickens and cows there are. Explain the reasoning.
+
+
+[Calibration sample 1-21]
+Three boxes are labeled apples, oranges, and mixed. Every label is wrong.
+You may take one fruit from one box. Explain how to correctly relabel
+all three boxes.
+
+
+[Calibration sample 1-22]
+A clock shows 3:15. What is the smaller angle between the hour hand
+and minute hand? Explain carefully.
+
+
+[Calibration sample 1-23]
+You have eight identical-looking balls and one is heavier.
+Using a balance scale only twice, explain how to find the heavier ball.
+
+
+[Calibration sample 1-24]
+A sequence contains 1, 1, 2, 3, 5, 8, 13. Identify the rule and
+calculate the next five values.
+
+
+[Calibration sample 1-25]
+Explain artificial intelligence to a ten-year-old using a simple
+real-world analogy. Avoid technical jargon where possible.
+
+
+[Calibration sample 1-26]
+Explain the difference between machine learning and traditional
+programming using a clear example.
+
+
+[Calibration sample 1-27]
+What is a neural network? Explain neurons, weights, training,
+and inference in simple but technically accurate language.
+
+
+[Calibration sample 1-28]
+Explain why training data quality can matter as much as training
+data quantity when building a language model.
+
+
+[Calibration sample 1-29]
+Describe the difference between an open-source model and an
+open-weight model. Explain why the distinction matters.
+
+
+[Calibration sample 1-30]
+اشرح مفهوم الذكاء الاصطناعي بطريقة بسيطة لطفل صغير، مع إعطاء أمثلة
+من الحياة اليومية وكيف يمكن للحاسوب أن يتعلم من البيانات.
+
+
+[Calibration sample 1-31]
+اشرح الفرق بين البرمجة التقليدية وتعلم الآلة، وقدم مثالا بسيطا
+يوضح كيف يتعلم النموذج من البيانات.
+
+
+[Calibration sample 1-32]
+حل المعادلة 3س + 7 = 31 واشرح جميع الخطوات بطريقة واضحة.
+
+
+[Calibration sample 1-33]
+ما هو النموذج اللغوي الكبير؟ اشرح كيف يتم تدريبه وكيف يستطيع
+توليد النصوص بعد التدريب.
+
+
+[Calibration sample 1-34]
+اشرح أهمية البيانات الجيدة عند تدريب نماذج الذكاء الاصطناعي
+ولماذا لا تعني كمية البيانات الكبيرة دائما جودة أفضل.
+
+
+[Calibration sample 1-35]
+شرح ليا شنو هو الذكاء الاصطناعي بطريقة بسيطة وبأمثلة من الحياة
+اليومية ديالنا فالمغرب.
+
+
+[Calibration sample 1-36]
+شنو الفرق بين البرمجة العادية وتعلم الآلة؟ شرحها ليا بطريقة سهلة
+وبمثال واضح.
+
+
+[Calibration sample 1-37]
+إلى كان عندنا 3x + 7 = 31، شحال كتساوي x؟ وشرح ليا المراحل وحدة بوحدة.
+
+
+[Calibration sample 1-38]
+شرح ليا شنو هو الموديل ديال اللغة الكبير وكيفاش كيتعلم من الداتا.
+
+
+[Calibration sample 1-39]
+علاش الداتا النظيفة والمزيانة مهمة فالتدريب ديال موديلات الذكاء
+الاصطناعي؟ وعلاش ماشي ديما الكمية الكبيرة هي الأفضل؟
+
+
+[Calibration sample 1-40]
+بغيت نبني تطبيق صغير بالبايثون كيجاوب على أسئلة المستخدمين.
+شنو هي المكونات اللي خاصني وكيفاش نقدر ننظم المشروع؟
+
+
+[Calibration sample 1-41]
+شرح ليا شنو هو API وكيفاش نقدر نستعملو فمشروع ديالي بالبايثون.
+
+
+[Calibration sample 1-42]
+عندي لائحة ديال الأرقام وبغيت نلقى أكبر مجموع ديال عناصر متتابعة
+فيها. شرح ليا كيفاش نخدم بخوارزمية Kadane.
+
+
+[Calibration sample 1-43]
+شنو الفرق بين CPU و GPU؟ وعلاش الـGPU مهم بزاف فالتدريب ديال
+موديلات الذكاء الاصطناعي؟
+
+
+[Calibration sample 1-44]
+شرح ليا علاش خاصنا نديرو validation للبيانات قبل ما ندربو موديل
+ديال machine learning.
+
+
+[Calibration sample 1-45]
+Large language models are neural networks trained to predict tokens.
+During training, the model processes sequences of text and adjusts its
+parameters so that its predictions become closer to the expected next
+tokens. Modern transformer architectures use attention mechanisms to
+connect information across different positions in a sequence. During
+inference, the trained parameters remain fixed while the model generates
+new tokens step by step.
+
+
+[Calibration sample 1-46]
+Quantization reduces the numerical precision used to represent model
+weights. A model stored in BF16 may require substantially more memory
+than the same model represented with four-bit or two-bit quantization.
+Lower-bit quantization can make large models easier to deploy, although
+aggressive quantization can reduce accuracy and may require techniques
+such as importance matrices to preserve the most important information.
+
+
+[Calibration sample 1-47]
+Software engineering requires careful handling of errors, interfaces,
+dependencies, testing, and maintainability. A program that works for one
+input is not necessarily robust. Good implementations consider edge
+cases, validate assumptions, use clear abstractions, and provide useful
+error messages when something goes wrong.
+
+
+[Calibration sample 1-48]
+When evaluating an AI model, it is useful to measure different
+capabilities separately. Coding benchmarks can test program synthesis,
+mathematical datasets can test symbolic and numerical reasoning, and
+language evaluations can measure instruction following and multilingual
+performance. A single benchmark score does not completely describe the
+capabilities of a model.
+
+
+[Calibration sample 2-0]
+Write a Python function that determines whether a string is a palindrome.
+The function should ignore spaces, punctuation, and capitalization.
+Explain the algorithm and its time complexity.
+
+
+[Calibration sample 2-1]
+Implement binary search in Python for a sorted list of integers.
+Return the index of the target or -1 when it is not present.
+Explain why binary search runs in logarithmic time.
+
+
+[Calibration sample 2-2]
+Write a Python function that counts the frequency of every word
+in a text. Normalize capitalization and ignore punctuation.
+Return the words ordered from most frequent to least frequent.
+
+
+[Calibration sample 2-3]
+Given a list of integers, find the maximum subarray sum using
+Kadane's algorithm. Explain why the algorithm works and give its
+time and space complexity.
+
+
+[Calibration sample 2-4]
+Implement a stack using two queues. Explain each operation and
+the tradeoff between the possible implementations.
+
+
+[Calibration sample 2-5]
+Write a function that checks whether two strings are anagrams.
+The solution should run in O(n) time using a frequency dictionary.
+
+
+[Calibration sample 2-6]
+Explain the difference between a process and a thread in an
+operating system. Give an example where multithreading is useful.
+
+
+[Calibration sample 2-7]
+Write a recursive function to calculate the factorial of an integer.
+Handle zero correctly and explain the base case.
+
+
+[Calibration sample 2-8]
+Explain what an API is and show a small Python example of sending
+a JSON request to an HTTP endpoint.
+
+
+[Calibration sample 2-9]
+Design a simple REST API for a chat application. Describe endpoints
+for creating conversations, sending messages, and retrieving history.
+
+
+[Calibration sample 2-10]
+Solve the equation 3x + 7 = 31. Show every algebraic step.
+
+
+[Calibration sample 2-11]
+A rectangle has a length of 12 centimeters and a width of 7 centimeters.
+Calculate its area and perimeter.
+
+
+[Calibration sample 2-12]
+If a train travels 240 kilometers in 3 hours, calculate its average
+speed in kilometers per hour.
+
+
+[Calibration sample 2-13]
+A number is multiplied by 4 and then increased by 9 to produce 37.
+Find the original number.
+
+
+[Calibration sample 2-14]
+There are 5 red balls and 3 blue balls in a bag. What is the probability
+of randomly selecting a blue ball? Explain your reasoning.
+
+
+[Calibration sample 2-15]
+A shop gives a 20 percent discount on a product priced at 250 dirhams.
+Calculate the discount and the final price.
+
+
+[Calibration sample 2-16]
+If x + y = 18 and x - y = 6, solve for x and y using a system of equations.
+
+
+[Calibration sample 2-17]
+Explain why the square root of 2 is irrational using a proof by contradiction.
+
+
+[Calibration sample 2-18]
+A sequence begins 2, 4, 8, 16, 32. Describe the pattern and calculate
+the next three terms.
+
+
+[Calibration sample 2-19]
+A triangle has a base of 10 centimeters and a height of 8 centimeters.
+Calculate its area and explain the formula.
+
+
+[Calibration sample 2-20]
+A farmer has chickens and cows. There are 20 animals and 56 legs.
+Determine how many chickens and cows there are. Explain the reasoning.
+
+
+[Calibration sample 2-21]
+Three boxes are labeled apples, oranges, and mixed. Every label is wrong.
+You may take one fruit from one box. Explain how to correctly relabel
+all three boxes.
+
+
+[Calibration sample 2-22]
+A clock shows 3:15. What is the smaller angle between the hour hand
+and minute hand? Explain carefully.
+
+
+[Calibration sample 2-23]
+You have eight identical-looking balls and one is heavier.
+Using a balance scale only twice, explain how to find the heavier ball.
+
+
+[Calibration sample 2-24]
+A sequence contains 1, 1, 2, 3, 5, 8, 13. Identify the rule and
+calculate the next five values.
+
+
+[Calibration sample 2-25]
+Explain artificial intelligence to a ten-year-old using a simple
+real-world analogy. Avoid technical jargon where possible.
+
+
+[Calibration sample 2-26]
+Explain the difference between machine learning and traditional
+programming using a clear example.
+
+
+[Calibration sample 2-27]
+What is a neural network? Explain neurons, weights, training,
+and inference in simple but technically accurate language.
+
+
+[Calibration sample 2-28]
+Explain why training data quality can matter as much as training
+data quantity when building a language model.
+
+
+[Calibration sample 2-29]
+Describe the difference between an open-source model and an
+open-weight model. Explain why the distinction matters.
+
+
+[Calibration sample 2-30]
+اشرح مفهوم الذكاء الاصطناعي بطريقة بسيطة لطفل صغير، مع إعطاء أمثلة
+من الحياة اليومية وكيف يمكن للحاسوب أن يتعلم من البيانات.
+
+
+[Calibration sample 2-31]
+اشرح الفرق بين البرمجة التقليدية وتعلم الآلة، وقدم مثالا بسيطا
+يوضح كيف يتعلم النموذج من البيانات.
+
+
+[Calibration sample 2-32]
+حل المعادلة 3س + 7 = 31 واشرح جميع الخطوات بطريقة واضحة.
+
+
+[Calibration sample 2-33]
+ما هو النموذج اللغوي الكبير؟ اشرح كيف يتم تدريبه وكيف يستطيع
+توليد النصوص بعد التدريب.
+
+
+[Calibration sample 2-34]
+اشرح أهمية البيانات الجيدة عند تدريب نماذج الذكاء الاصطناعي
+ولماذا لا تعني كمية البيانات الكبيرة دائما جودة أفضل.
+
+
+[Calibration sample 2-35]
+شرح ليا شنو هو الذكاء الاصطناعي بطريقة بسيطة وبأمثلة من الحياة
+اليومية ديالنا فالمغرب.
+
+
+[Calibration sample 2-36]
+شنو الفرق بين البرمجة العادية وتعلم الآلة؟ شرحها ليا بطريقة سهلة
+وبمثال واضح.
+
+
+[Calibration sample 2-37]
+إلى كان عندنا 3x + 7 = 31، شحال كتساوي x؟ وشرح ليا المراحل وحدة بوحدة.
+
+
+[Calibration sample 2-38]
+شرح ليا شنو هو الموديل ديال اللغة الكبير وكيفاش كيتعلم من الداتا.
+
+
+[Calibration sample 2-39]
+علاش الداتا النظيفة والمزيانة مهمة فالتدريب ديال موديلات الذكاء
+الاصطناعي؟ وعلاش ماشي ديما الكمية الكبيرة هي الأفضل؟
+
+
+[Calibration sample 2-40]
+بغيت نبني تطبيق صغير بالبايثون كيجاوب على أسئلة المستخدمين.
+شنو هي المكونات اللي خاصني وكيفاش نقدر ننظم المشروع؟
+
+
+[Calibration sample 2-41]
+شرح ليا شنو هو API وكيفاش نقدر نستعملو فمشروع ديالي بالبايثون.
+
+
+[Calibration sample 2-42]
+عندي لائحة ديال الأرقام وبغيت نلقى أكبر مجموع ديال عناصر متتابعة
+فيها. شرح ليا كيفاش نخدم بخوارزمية Kadane.
+
+
+[Calibration sample 2-43]
+شنو الفرق بين CPU و GPU؟ وعلاش الـGPU مهم بزاف فالتدريب ديال
+موديلات الذكاء الاصطناعي؟
+
+
+[Calibration sample 2-44]
+شرح ليا علاش خاصنا نديرو validation للبيانات قبل ما ندربو موديل
+ديال machine learning.
+
+
+[Calibration sample 2-45]
+Large language models are neural networks trained to predict tokens.
+During training, the model processes sequences of text and adjusts its
+parameters so that its predictions become closer to the expected next
+tokens. Modern transformer architectures use attention mechanisms to
+connect information across different positions in a sequence. During
+inference, the trained parameters remain fixed while the model generates
+new tokens step by step.
+
+
+[Calibration sample 2-46]
+Quantization reduces the numerical precision used to represent model
+weights. A model stored in BF16 may require substantially more memory
+than the same model represented with four-bit or two-bit quantization.
+Lower-bit quantization can make large models easier to deploy, although
+aggressive quantization can reduce accuracy and may require techniques
+such as importance matrices to preserve the most important information.
+
+
+[Calibration sample 2-47]
+Software engineering requires careful handling of errors, interfaces,
+dependencies, testing, and maintainability. A program that works for one
+input is not necessarily robust. Good implementations consider edge
+cases, validate assumptions, use clear abstractions, and provide useful
+error messages when something goes wrong.
+
+
+[Calibration sample 2-48]
+When evaluating an AI model, it is useful to measure different
+capabilities separately. Coding benchmarks can test program synthesis,
+mathematical datasets can test symbolic and numerical reasoning, and
+language evaluations can measure instruction following and multilingual
+performance. A single benchmark score does not completely describe the
+capabilities of a model.
+
+
+[Calibration sample 3-0]
+Write a Python function that determines whether a string is a palindrome.
+The function should ignore spaces, punctuation, and capitalization.
+Explain the algorithm and its time complexity.
+
+
+[Calibration sample 3-1]
+Implement binary search in Python for a sorted list of integers.
+Return the index of the target or -1 when it is not present.
+Explain why binary search runs in logarithmic time.
+
+
+[Calibration sample 3-2]
+Write a Python function that counts the frequency of every word
+in a text. Normalize capitalization and ignore punctuation.
+Return the words ordered from most frequent to least frequent.
+
+
+[Calibration sample 3-3]
+Given a list of integers, find the maximum subarray sum using
+Kadane's algorithm. Explain why the algorithm works and give its
+time and space complexity.
+
+
+[Calibration sample 3-4]
+Implement a stack using two queues. Explain each operation and
+the tradeoff between the possible implementations.
+
+
+[Calibration sample 3-5]
+Write a function that checks whether two strings are anagrams.
+The solution should run in O(n) time using a frequency dictionary.
+
+
+[Calibration sample 3-6]
+Explain the difference between a process and a thread in an
+operating system. Give an example where multithreading is useful.
+
+
+[Calibration sample 3-7]
+Write a recursive function to calculate the factorial of an integer.
+Handle zero correctly and explain the base case.
+
+
+[Calibration sample 3-8]
+Explain what an API is and show a small Python example of sending
+a JSON request to an HTTP endpoint.
+
+
+[Calibration sample 3-9]
+Design a simple REST API for a chat application. Describe endpoints
+for creating conversations, sending messages, and retrieving history.
+
+
+[Calibration sample 3-10]
+Solve the equation 3x + 7 = 31. Show every algebraic step.
+
+
+[Calibration sample 3-11]
+A rectangle has a length of 12 centimeters and a width of 7 centimeters.
+Calculate its area and perimeter.
+
+
+[Calibration sample 3-12]
+If a train travels 240 kilometers in 3 hours, calculate its average
+speed in kilometers per hour.
+
+
+[Calibration sample 3-13]
+A number is multiplied by 4 and then increased by 9 to produce 37.
+Find the original number.
+
+
+[Calibration sample 3-14]
+There are 5 red balls and 3 blue balls in a bag. What is the probability
+of randomly selecting a blue ball? Explain your reasoning.
+
+
+[Calibration sample 3-15]
+A shop gives a 20 percent discount on a product priced at 250 dirhams.
+Calculate the discount and the final price.
+
+
+[Calibration sample 3-16]
+If x + y = 18 and x - y = 6, solve for x and y using a system of equations.
+
+
+[Calibration sample 3-17]
+Explain why the square root of 2 is irrational using a proof by contradiction.
+
+
+[Calibration sample 3-18]
+A sequence begins 2, 4, 8, 16, 32. Describe the pattern and calculate
+the next three terms.
+
+
+[Calibration sample 3-19]
+A triangle has a base of 10 centimeters and a height of 8 centimeters.
+Calculate its area and explain the formula.
+
+
+[Calibration sample 3-20]
+A farmer has chickens and cows. There are 20 animals and 56 legs.
+Determine how many chickens and cows there are. Explain the reasoning.
+
+
+[Calibration sample 3-21]
+Three boxes are labeled apples, oranges, and mixed. Every label is wrong.
+You may take one fruit from one box. Explain how to correctly relabel
+all three boxes.
+
+
+[Calibration sample 3-22]
+A clock shows 3:15. What is the smaller angle between the hour hand
+and minute hand? Explain carefully.
+
+
+[Calibration sample 3-23]
+You have eight identical-looking balls and one is heavier.
+Using a balance scale only twice, explain how to find the heavier ball.
+
+
+[Calibration sample 3-24]
+A sequence contains 1, 1, 2, 3, 5, 8, 13. Identify the rule and
+calculate the next five values.
+
+
+[Calibration sample 3-25]
+Explain artificial intelligence to a ten-year-old using a simple
+real-world analogy. Avoid technical jargon where possible.
+
+
+[Calibration sample 3-26]
+Explain the difference between machine learning and traditional
+programming using a clear example.
+
+
+[Calibration sample 3-27]
+What is a neural network? Explain neurons, weights, training,
+and inference in simple but technically accurate language.
+
+
+[Calibration sample 3-28]
+Explain why training data quality can matter as much as training
+data quantity when building a language model.
+
+
+[Calibration sample 3-29]
+Describe the difference between an open-source model and an
+open-weight model. Explain why the distinction matters.
+
+
+[Calibration sample 3-30]
+اشرح مفهوم الذكاء الاصطناعي بطريقة بسيطة لطفل صغير، مع إعطاء أمثلة
+من الحياة اليومية وكيف يمكن للحاسوب أن يتعلم من البيانات.
+
+
+[Calibration sample 3-31]
+اشرح الفرق بين البرمجة التقليدية وتعلم الآلة، وقدم مثالا بسيطا
+يوضح كيف يتعلم النموذج من البيانات.
+
+
+[Calibration sample 3-32]
+حل المعادلة 3س + 7 = 31 واشرح جميع الخطوات بطريقة واضحة.
+
+
+[Calibration sample 3-33]
+ما هو النموذج اللغوي الكبير؟ اشرح كيف يتم تدريبه وكيف يستطيع
+توليد النصوص بعد التدريب.
+
+
+[Calibration sample 3-34]
+اشرح أهمية البيانات الجيدة عند تدريب نماذج الذكاء الاصطناعي
+ولماذا لا تعني كمية البيانات الكبيرة دائما جودة أفضل.
+
+
+[Calibration sample 3-35]
+شرح ليا شنو هو الذكاء الاصطناعي بطريقة بسيطة وبأمثلة من الحياة
+اليومية ديالنا فالمغرب.
+
+
+[Calibration sample 3-36]
+شنو الفرق بين البرمجة العادية وتعلم الآلة؟ شرحها ليا بطريقة سهلة
+وبمثال واضح.
+
+
+[Calibration sample 3-37]
+إلى كان عندنا 3x + 7 = 31، شحال كتساوي x؟ وشرح ليا المراحل وحدة بوحدة.
+
+
+[Calibration sample 3-38]
+شرح ليا شنو هو الموديل ديال اللغة الكبير وكيفاش كيتعلم من الداتا.
+
+
+[Calibration sample 3-39]
+علاش الداتا النظيفة والمزيانة مهمة فالتدريب ديال موديلات الذكاء
+الاصطناعي؟ وعلاش ماشي ديما الكمية الكبيرة هي الأفضل؟
+
+
+[Calibration sample 3-40]
+بغيت نبني تطبيق صغير بالبايثون كيجاوب على أسئلة المستخدمين.
+شنو هي المكونات اللي خاصني وكيفاش نقدر ننظم المشروع؟
+
+
+[Calibration sample 3-41]
+شرح ليا شنو هو API وكيفاش نقدر نستعملو فمشروع ديالي بالبايثون.
+
+
+[Calibration sample 3-42]
+عندي لائحة ديال الأرقام وبغيت نلقى أكبر مجموع ديال عناصر متتابعة
+فيها. شرح ليا كيفاش نخدم بخوارزمية Kadane.
+
+
+[Calibration sample 3-43]
+شنو الفرق بين CPU و GPU؟ وعلاش الـGPU مهم بزاف فالتدريب ديال
+موديلات الذكاء الاصطناعي؟
+
+
+[Calibration sample 3-44]
+شرح ليا علاش خاصنا نديرو validation للبيانات قبل ما ندربو موديل
+ديال machine learning.
+
+
+[Calibration sample 3-45]
+Large language models are neural networks trained to predict tokens.
+During training, the model processes sequences of text and adjusts its
+parameters so that its predictions become closer to the expected next
+tokens. Modern transformer architectures use attention mechanisms to
+connect information across different positions in a sequence. During
+inference, the trained parameters remain fixed while the model generates
+new tokens step by step.
+
+
+[Calibration sample 3-46]
+Quantization reduces the numerical precision used to represent model
+weights. A model stored in BF16 may require substantially more memory
+than the same model represented with four-bit or two-bit quantization.
+Lower-bit quantization can make large models easier to deploy, although
+aggressive quantization can reduce accuracy and may require techniques
+such as importance matrices to preserve the most important information.
+
+
+[Calibration sample 3-47]
+Software engineering requires careful handling of errors, interfaces,
+dependencies, testing, and maintainability. A program that works for one
+input is not necessarily robust. Good implementations consider edge
+cases, validate assumptions, use clear abstractions, and provide useful
+error messages when something goes wrong.
+
+
+[Calibration sample 3-48]
+When evaluating an AI model, it is useful to measure different
+capabilities separately. Coding benchmarks can test program synthesis,
+mathematical datasets can test symbolic and numerical reasoning, and
+language evaluations can measure instruction following and multilingual
+performance. A single benchmark score does not completely describe the
+capabilities of a model.
+
+
+[Calibration sample 4-0]
+Write a Python function that determines whether a string is a palindrome.
+The function should ignore spaces, punctuation, and capitalization.
+Explain the algorithm and its time complexity.
+
+
+[Calibration sample 4-1]
+Implement binary search in Python for a sorted list of integers.
+Return the index of the target or -1 when it is not present.
+Explain why binary search runs in logarithmic time.
+
+
+[Calibration sample 4-2]
+Write a Python function that counts the frequency of every word
+in a text. Normalize capitalization and ignore punctuation.
+Return the words ordered from most frequent to least frequent.
+
+
+[Calibration sample 4-3]
+Given a list of integers, find the maximum subarray sum using
+Kadane's algorithm. Explain why the algorithm works and give its
+time and space complexity.
+
+
+[Calibration sample 4-4]
+Implement a stack using two queues. Explain each operation and
+the tradeoff between the possible implementations.
+
+
+[Calibration sample 4-5]
+Write a function that checks whether two strings are anagrams.
+The solution should run in O(n) time using a frequency dictionary.
+
+
+[Calibration sample 4-6]
+Explain the difference between a process and a thread in an
+operating system. Give an example where multithreading is useful.
+
+
+[Calibration sample 4-7]
+Write a recursive function to calculate the factorial of an integer.
+Handle zero correctly and explain the base case.
+
+
+[Calibration sample 4-8]
+Explain what an API is and show a small Python example of sending
+a JSON request to an HTTP endpoint.
+
+
+[Calibration sample 4-9]
+Design a simple REST API for a chat application. Describe endpoints
+for creating conversations, sending messages, and retrieving history.
+
+
+[Calibration sample 4-10]
+Solve the equation 3x + 7 = 31. Show every algebraic step.
+
+
+[Calibration sample 4-11]
+A rectangle has a length of 12 centimeters and a width of 7 centimeters.
+Calculate its area and perimeter.
+
+
+[Calibration sample 4-12]
+If a train travels 240 kilometers in 3 hours, calculate its average
+speed in kilometers per hour.
+
+
+[Calibration sample 4-13]
+A number is multiplied by 4 and then increased by 9 to produce 37.
+Find the original number.
+
+
+[Calibration sample 4-14]
+There are 5 red balls and 3 blue balls in a bag. What is the probability
+of randomly selecting a blue ball? Explain your reasoning.
+
+
+[Calibration sample 4-15]
+A shop gives a 20 percent discount on a product priced at 250 dirhams.
+Calculate the discount and the final price.
+
+
+[Calibration sample 4-16]
+If x + y = 18 and x - y = 6, solve for x and y using a system of equations.
+
+
+[Calibration sample 4-17]
+Explain why the square root of 2 is irrational using a proof by contradiction.
+
+
+[Calibration sample 4-18]
+A sequence begins 2, 4, 8, 16, 32. Describe the pattern and calculate
+the next three terms.
+
+
+[Calibration sample 4-19]
+A triangle has a base of 10 centimeters and a height of 8 centimeters.
+Calculate its area and explain the formula.
+
+
+[Calibration sample 4-20]
+A farmer has chickens and cows. There are 20 animals and 56 legs.
+Determine how many chickens and cows there are. Explain the reasoning.
+
+
+[Calibration sample 4-21]
+Three boxes are labeled apples, oranges, and mixed. Every label is wrong.
+You may take one fruit from one box. Explain how to correctly relabel
+all three boxes.
+
+
+[Calibration sample 4-22]
+A clock shows 3:15. What is the smaller angle between the hour hand
+and minute hand? Explain carefully.
+
+
+[Calibration sample 4-23]
+You have eight identical-looking balls and one is heavier.
+Using a balance scale only twice, explain how to find the heavier ball.
+
+
+[Calibration sample 4-24]
+A sequence contains 1, 1, 2, 3, 5, 8, 13. Identify the rule and
+calculate the next five values.
+
+
+[Calibration sample 4-25]
+Explain artificial intelligence to a ten-year-old using a simple
+real-world analogy. Avoid technical jargon where possible.
+
+
+[Calibration sample 4-26]
+Explain the difference between machine learning and traditional
+programming using a clear example.
+
+
+[Calibration sample 4-27]
+What is a neural network? Explain neurons, weights, training,
+and inference in simple but technically accurate language.
+
+
+[Calibration sample 4-28]
+Explain why training data quality can matter as much as training
+data quantity when building a language model.
+
+
+[Calibration sample 4-29]
+Describe the difference between an open-source model and an
+open-weight model. Explain why the distinction matters.
+
+
+[Calibration sample 4-30]
+اشرح مفهوم الذكاء الاصطناعي بطريقة بسيطة لطفل صغير، مع إعطاء أمثلة
+من الحياة اليومية وكيف يمكن للحاسوب أن يتعلم من البيانات.
+
+
+[Calibration sample 4-31]
+اشرح الفرق بين البرمجة التقليدية وتعلم الآلة، وقدم مثالا بسيطا
+يوضح كيف يتعلم النموذج من البيانات.
+
+
+[Calibration sample 4-32]
+حل المعادلة 3س + 7 = 31 واشرح جميع الخطوات بطريقة واضحة.
+
+
+[Calibration sample 4-33]
+ما هو النموذج اللغوي الكبير؟ اشرح كيف يتم تدريبه وكيف يستطيع
+توليد النصوص بعد التدريب.
+
+
+[Calibration sample 4-34]
+اشرح أهمية البيانات الجيدة عند تدريب نماذج الذكاء الاصطناعي
+ولماذا لا تعني كمية البيانات الكبيرة دائما جودة أفضل.
+
+
+[Calibration sample 4-35]
+شرح ليا شنو هو الذكاء الاصطناعي بطريقة بسيطة وبأمثلة من الحياة
+اليومية ديالنا فالمغرب.
+
+
+[Calibration sample 4-36]
+شنو الفرق بين البرمجة العادية وتعلم الآلة؟ شرحها ليا بطريقة سهلة
+وبمثال واضح.
+
+
+[Calibration sample 4-37]
+إلى كان عندنا 3x + 7 = 31، شحال كتساوي x؟ وشرح ليا المراحل وحدة بوحدة.
+
+
+[Calibration sample 4-38]
+شرح ليا شنو هو الموديل ديال اللغة الكبير وكيفاش كيتعلم من الداتا.
+
+
+[Calibration sample 4-39]
+علاش الداتا النظيفة والمزيانة مهمة فالتدريب ديال موديلات الذكاء
+الاصطناعي؟ وعلاش ماشي ديما الكمية الكبيرة هي الأفضل؟
+
+
+[Calibration sample 4-40]
+بغيت نبني تطبيق صغير بالبايثون كيجاوب على أسئلة المستخدمين.
+شنو هي المكونات اللي خاصني وكيفاش نقدر ننظم المشروع؟
+
+
+[Calibration sample 4-41]
+شرح ليا شنو هو API وكيفاش نقدر نستعملو فمشروع ديالي بالبايثون.
+
+
+[Calibration sample 4-42]
+عندي لائحة ديال الأرقام وبغيت نلقى أكبر مجموع ديال عناصر متتابعة
+فيها. شرح ليا كيفاش نخدم بخوارزمية Kadane.
+
+
+[Calibration sample 4-43]
+شنو الفرق بين CPU و GPU؟ وعلاش الـGPU مهم بزاف فالتدريب ديال
+موديلات الذكاء الاصطناعي؟
+
+
+[Calibration sample 4-44]
+شرح ليا علاش خاصنا نديرو validation للبيانات قبل ما ندربو موديل
+ديال machine learning.
+
+
+[Calibration sample 4-45]
+Large language models are neural networks trained to predict tokens.
+During training, the model processes sequences of text and adjusts its
+parameters so that its predictions become closer to the expected next
+tokens. Modern transformer architectures use attention mechanisms to
+connect information across different positions in a sequence. During
+inference, the trained parameters remain fixed while the model generates
+new tokens step by step.
+
+
+[Calibration sample 4-46]
+Quantization reduces the numerical precision used to represent model
+weights. A model stored in BF16 may require substantially more memory
+than the same model represented with four-bit or two-bit quantization.
+Lower-bit quantization can make large models easier to deploy, although
+aggressive quantization can reduce accuracy and may require techniques
+such as importance matrices to preserve the most important information.
+
+
+[Calibration sample 4-47]
+Software engineering requires careful handling of errors, interfaces,
+dependencies, testing, and maintainability. A program that works for one
+input is not necessarily robust. Good implementations consider edge
+cases, validate assumptions, use clear abstractions, and provide useful
+error messages when something goes wrong.
+
+
+[Calibration sample 4-48]
+When evaluating an AI model, it is useful to measure different
+capabilities separately. Coding benchmarks can test program synthesis,
+mathematical datasets can test symbolic and numerical reasoning, and
+language evaluations can measure instruction following and multilingual
+performance. A single benchmark score does not completely describe the
+capabilities of a model.
+
+
+[Calibration sample 5-0]
+Write a Python function that determines whether a string is a palindrome.
+The function should ignore spaces, punctuation, and capitalization.
+Explain the algorithm and its time complexity.
+
+
+[Calibration sample 5-1]
+Implement binary search in Python for a sorted list of integers.
+Return the index of the target or -1 when it is not present.
+Explain why binary search runs in logarithmic time.
+
+
+[Calibration sample 5-2]
+Write a Python function that counts the frequency of every word
+in a text. Normalize capitalization and ignore punctuation.
+Return the words ordered from most frequent to least frequent.
+
+
+[Calibration sample 5-3]
+Given a list of integers, find the maximum subarray sum using
+Kadane's algorithm. Explain why the algorithm works and give its
+time and space complexity.
+
+
+[Calibration sample 5-4]
+Implement a stack using two queues. Explain each operation and
+the tradeoff between the possible implementations.
+
+
+[Calibration sample 5-5]
+Write a function that checks whether two strings are anagrams.
+The solution should run in O(n) time using a frequency dictionary.
+
+
+[Calibration sample 5-6]
+Explain the difference between a process and a thread in an
+operating system. Give an example where multithreading is useful.
+
+
+[Calibration sample 5-7]
+Write a recursive function to calculate the factorial of an integer.
+Handle zero correctly and explain the base case.
+
+
+[Calibration sample 5-8]
+Explain what an API is and show a small Python example of sending
+a JSON request to an HTTP endpoint.
+
+
+[Calibration sample 5-9]
+Design a simple REST API for a chat application. Describe endpoints
+for creating conversations, sending messages, and retrieving history.
+
+
+[Calibration sample 5-10]
+Solve the equation 3x + 7 = 31. Show every algebraic step.
+
+
+[Calibration sample 5-11]
+A rectangle has a length of 12 centimeters and a width of 7 centimeters.
+Calculate its area and perimeter.
+
+
+[Calibration sample 5-12]
+If a train travels 240 kilometers in 3 hours, calculate its average
+speed in kilometers per hour.
+
+
+[Calibration sample 5-13]
+A number is multiplied by 4 and then increased by 9 to produce 37.
+Find the original number.
+
+
+[Calibration sample 5-14]
+There are 5 red balls and 3 blue balls in a bag. What is the probability
+of randomly selecting a blue ball? Explain your reasoning.
+
+
+[Calibration sample 5-15]
+A shop gives a 20 percent discount on a product priced at 250 dirhams.
+Calculate the discount and the final price.
+
+
+[Calibration sample 5-16]
+If x + y = 18 and x - y = 6, solve for x and y using a system of equations.
+
+
+[Calibration sample 5-17]
+Explain why the square root of 2 is irrational using a proof by contradiction.
+
+
+[Calibration sample 5-18]
+A sequence begins 2, 4, 8, 16, 32. Describe the pattern and calculate
+the next three terms.
+
+
+[Calibration sample 5-19]
+A triangle has a base of 10 centimeters and a height of 8 centimeters.
+Calculate its area and explain the formula.
+
+
+[Calibration sample 5-20]
+A farmer has chickens and cows. There are 20 animals and 56 legs.
+Determine how many chickens and cows there are. Explain the reasoning.
+
+
+[Calibration sample 5-21]
+Three boxes are labeled apples, oranges, and mixed. Every label is wrong.
+You may take one fruit from one box. Explain how to correctly relabel
+all three boxes.
+
+
+[Calibration sample 5-22]
+A clock shows 3:15. What is the smaller angle between the hour hand
+and minute hand? Explain carefully.
+
+
+[Calibration sample 5-23]
+You have eight identical-looking balls and one is heavier.
+Using a balance scale only twice, explain how to find the heavier ball.
+
+
+[Calibration sample 5-24]
+A sequence contains 1, 1, 2, 3, 5, 8, 13. Identify the rule and
+calculate the next five values.
+
+
+[Calibration sample 5-25]
+Explain artificial intelligence to a ten-year-old using a simple
+real-world analogy. Avoid technical jargon where possible.
+
+
+[Calibration sample 5-26]
+Explain the difference between machine learning and traditional
+programming using a clear example.
+
+
+[Calibration sample 5-27]
+What is a neural network? Explain neurons, weights, training,
+and inference in simple but technically accurate language.
+
+
+[Calibration sample 5-28]
+Explain why training data quality can matter as much as training
+data quantity when building a language model.
+
+
+[Calibration sample 5-29]
+Describe the difference between an open-source model and an
+open-weight model. Explain why the distinction matters.
+
+
+[Calibration sample 5-30]
+اشرح مفهوم الذكاء الاصطناعي بطريقة بسيطة لطفل صغير، مع إعطاء أمثلة
+من الحياة اليومية وكيف يمكن للحاسوب أن يتعلم من البيانات.
+
+
+[Calibration sample 5-31]
+اشرح الفرق بين البرمجة التقليدية وتعلم الآلة، وقدم مثالا بسيطا
+يوضح كيف يتعلم النموذج من البيانات.
+
+
+[Calibration sample 5-32]
+حل المعادلة 3س + 7 = 31 واشرح جميع الخطوات بطريقة واضحة.
+
+
+[Calibration sample 5-33]
+ما هو النموذج اللغوي الكبير؟ اشرح كيف يتم تدريبه وكيف يستطيع
+توليد النصوص بعد التدريب.
+
+
+[Calibration sample 5-34]
+اشرح أهمية البيانات الجيدة عند تدريب نماذج الذكاء الاصطناعي
+ولماذا لا تعني كمية البيانات الكبيرة دائما جودة أفضل.
+
+
+[Calibration sample 5-35]
+شرح ليا شنو هو الذكاء الاصطناعي بطريقة بسيطة وبأمثلة من الحياة
+اليومية ديالنا فالمغرب.
+
+
+[Calibration sample 5-36]
+شنو الفرق بين البرمجة العادية وتعلم الآلة؟ شرحها ليا بطريقة سهلة
+وبمثال واضح.
+
+
+[Calibration sample 5-37]
+إلى كان عندنا 3x + 7 = 31، شحال كتساوي x؟ وشرح ليا المراحل وحدة بوحدة.
+
+
+[Calibration sample 5-38]
+شرح ليا شنو هو الموديل ديال اللغة الكبير وكيفاش كيتعلم من الداتا.
+
+
+[Calibration sample 5-39]
+علاش الداتا النظيفة والمزيانة مهمة فالتدريب ديال موديلات الذكاء
+الاصطناعي؟ وعلاش ماشي ديما الكمية الكبيرة هي الأفضل؟
+
+
+[Calibration sample 5-40]
+بغيت نبني تطبيق صغير بالبايثون كيجاوب على أسئلة المستخدمين.
+شنو هي المكونات اللي خاصني وكيفاش نقدر ننظم المشروع؟
+
+
+[Calibration sample 5-41]
+شرح ليا شنو هو API وكيفاش نقدر نستعملو فمشروع ديالي بالبايثون.
+
+
+[Calibration sample 5-42]
+عندي لائحة ديال الأرقام وبغيت نلقى أكبر مجموع ديال عناصر متتابعة
+فيها. شرح ليا كيفاش نخدم بخوارزمية Kadane.
+
+
+[Calibration sample 5-43]
+شنو الفرق بين CPU و GPU؟ وعلاش الـGPU مهم بزاف فالتدريب ديال
+موديلات الذكاء الاصطناعي؟
+
+
+[Calibration sample 5-44]
+شرح ليا علاش خاصنا نديرو validation للبيانات قبل ما ندربو موديل
+ديال machine learning.
+
+
+[Calibration sample 5-45]
+Large language models are neural networks trained to predict tokens.
+During training, the model processes sequences of text and adjusts its
+parameters so that its predictions become closer to the expected next
+tokens. Modern transformer architectures use attention mechanisms to
+connect information across different positions in a sequence. During
+inference, the trained parameters remain fixed while the model generates
+new tokens step by step.
+
+
+[Calibration sample 5-46]
+Quantization reduces the numerical precision used to represent model
+weights. A model stored in BF16 may require substantially more memory
+than the same model represented with four-bit or two-bit quantization.
+Lower-bit quantization can make large models easier to deploy, although
+aggressive quantization can reduce accuracy and may require techniques
+such as importance matrices to preserve the most important information.
+
+
+[Calibration sample 5-47]
+Software engineering requires careful handling of errors, interfaces,
+dependencies, testing, and maintainability. A program that works for one
+input is not necessarily robust. Good implementations consider edge
+cases, validate assumptions, use clear abstractions, and provide useful
+error messages when something goes wrong.
+
+
+[Calibration sample 5-48]
+When evaluating an AI model, it is useful to measure different
+capabilities separately. Coding benchmarks can test program synthesis,
+mathematical datasets can test symbolic and numerical reasoning, and
+language evaluations can measure instruction following and multilingual
+performance. A single benchmark score does not completely describe the
+capabilities of a model.
+
+
+[Calibration sample 6-0]
+Write a Python function that determines whether a string is a palindrome.
+The function should ignore spaces, punctuation, and capitalization.
+Explain the algorithm and its time complexity.
+
+
+[Calibration sample 6-1]
+Implement binary search in Python for a sorted list of integers.
+Return the index of the target or -1 when it is not present.
+Explain why binary search runs in logarithmic time.
+
+
+[Calibration sample 6-2]
+Write a Python function that counts the frequency of every word
+in a text. Normalize capitalization and ignore punctuation.
+Return the words ordered from most frequent to least frequent.
+
+
+[Calibration sample 6-3]
+Given a list of integers, find the maximum subarray sum using
+Kadane's algorithm. Explain why the algorithm works and give its
+time and space complexity.
+
+
+[Calibration sample 6-4]
+Implement a stack using two queues. Explain each operation and
+the tradeoff between the possible implementations.
+
+
+[Calibration sample 6-5]
+Write a function that checks whether two strings are anagrams.
+The solution should run in O(n) time using a frequency dictionary.
+
+
+[Calibration sample 6-6]
+Explain the difference between a process and a thread in an
+operating system. Give an example where multithreading is useful.
+
+
+[Calibration sample 6-7]
+Write a recursive function to calculate the factorial of an integer.
+Handle zero correctly and explain the base case.
+
+
+[Calibration sample 6-8]
+Explain what an API is and show a small Python example of sending
+a JSON request to an HTTP endpoint.
+
+
+[Calibration sample 6-9]
+Design a simple REST API for a chat application. Describe endpoints
+for creating conversations, sending messages, and retrieving history.
+
+
+[Calibration sample 6-10]
+Solve the equation 3x + 7 = 31. Show every algebraic step.
+
+
+[Calibration sample 6-11]
+A rectangle has a length of 12 centimeters and a width of 7 centimeters.
+Calculate its area and perimeter.
+
+
+[Calibration sample 6-12]
+If a train travels 240 kilometers in 3 hours, calculate its average
+speed in kilometers per hour.
+
+
+[Calibration sample 6-13]
+A number is multiplied by 4 and then increased by 9 to produce 37.
+Find the original number.
+
+
+[Calibration sample 6-14]
+There are 5 red balls and 3 blue balls in a bag. What is the probability
+of randomly selecting a blue ball? Explain your reasoning.
+
+
+[Calibration sample 6-15]
+A shop gives a 20 percent discount on a product priced at 250 dirhams.
+Calculate the discount and the final price.
+
+
+[Calibration sample 6-16]
+If x + y = 18 and x - y = 6, solve for x and y using a system of equations.
+
+
+[Calibration sample 6-17]
+Explain why the square root of 2 is irrational using a proof by contradiction.
+
+
+[Calibration sample 6-18]
+A sequence begins 2, 4, 8, 16, 32. Describe the pattern and calculate
+the next three terms.
+
+
+[Calibration sample 6-19]
+A triangle has a base of 10 centimeters and a height of 8 centimeters.
+Calculate its area and explain the formula.
+
+
+[Calibration sample 6-20]
+A farmer has chickens and cows. There are 20 animals and 56 legs.
+Determine how many chickens and cows there are. Explain the reasoning.
+
+
+[Calibration sample 6-21]
+Three boxes are labeled apples, oranges, and mixed. Every label is wrong.
+You may take one fruit from one box. Explain how to correctly relabel
+all three boxes.
+
+
+[Calibration sample 6-22]
+A clock shows 3:15. What is the smaller angle between the hour hand
+and minute hand? Explain carefully.
+
+
+[Calibration sample 6-23]
+You have eight identical-looking balls and one is heavier.
+Using a balance scale only twice, explain how to find the heavier ball.
+
+
+[Calibration sample 6-24]
+A sequence contains 1, 1, 2, 3, 5, 8, 13. Identify the rule and
+calculate the next five values.
+
+
+[Calibration sample 6-25]
+Explain artificial intelligence to a ten-year-old using a simple
+real-world analogy. Avoid technical jargon where possible.
+
+
+[Calibration sample 6-26]
+Explain the difference between machine learning and traditional
+programming using a clear example.
+
+
+[Calibration sample 6-27]
+What is a neural network? Explain neurons, weights, training,
+and inference in simple but technically accurate language.
+
+
+[Calibration sample 6-28]
+Explain why training data quality can matter as much as training
+data quantity when building a language model.
+
+
+[Calibration sample 6-29]
+Describe the difference between an open-source model and an
+open-weight model. Explain why the distinction matters.
+
+
+[Calibration sample 6-30]
+اشرح مفهوم الذكاء الاصطناعي بطريقة بسيطة لطفل صغير، مع إعطاء أمثلة
+من الحياة اليومية وكيف يمكن للحاسوب أن يتعلم من البيانات.
+
+
+[Calibration sample 6-31]
+اشرح الفرق بين البرمجة التقليدية وتعلم الآلة، وقدم مثالا بسيطا
+يوضح كيف يتعلم النموذج من البيانات.
+
+
+[Calibration sample 6-32]
+حل المعادلة 3س + 7 = 31 واشرح جميع الخطوات بطريقة واضحة.
+
+
+[Calibration sample 6-33]
+ما هو النموذج اللغوي الكبير؟ اشرح كيف يتم تدريبه وكيف يستطيع
+توليد النصوص بعد التدريب.
+
+
+[Calibration sample 6-34]
+اشرح أهمية البيانات الجيدة عند تدريب نماذج الذكاء الاصطناعي
+ولماذا لا تعني كمية البيانات الكبيرة دائما جودة أفضل.
+
+
+[Calibration sample 6-35]
+شرح ليا شنو هو الذكاء الاصطناعي بطريقة بسيطة وبأمثلة من الحياة
+اليومية ديالنا فالمغرب.
+
+
+[Calibration sample 6-36]
+شنو الفرق بين البرمجة العادية وتعلم الآلة؟ شرحها ليا بطريقة سهلة
+وبمثال واضح.
+
+
+[Calibration sample 6-37]
+إلى كان عندنا 3x + 7 = 31، شحال كتساوي x؟ وشرح ليا المراحل وحدة بوحدة.
+
+
+[Calibration sample 6-38]
+شرح ليا شنو هو الموديل ديال اللغة الكبير وكيفاش كيتعلم من الداتا.
+
+
+[Calibration sample 6-39]
+علاش الداتا النظيفة والمزيانة مهمة فالتدريب ديال موديلات الذكاء
+الاصطناعي؟ وعلاش ماشي ديما الكمية الكبيرة هي الأفضل؟
+
+
+[Calibration sample 6-40]
+بغيت نبني تطبيق صغير بالبايثون كيجاوب على أسئلة المستخدمين.
+شنو هي المكونات اللي خاصني وكيفاش نقدر ننظم المشروع؟
+
+
+[Calibration sample 6-41]
+شرح ليا شنو هو API وكيفاش نقدر نستعملو فمشروع ديالي بالبايثون.
+
+
+[Calibration sample 6-42]
+عندي لائحة ديال الأرقام وبغيت نلقى أكبر مجموع ديال عناصر متتابعة
+فيها. شرح ليا كيفاش نخدم بخوارزمية Kadane.
+
+
+[Calibration sample 6-43]
+شنو الفرق بين CPU و GPU؟ وعلاش الـGPU مهم بزاف فالتدريب ديال
+موديلات الذكاء الاصطناعي؟
+
+
+[Calibration sample 6-44]
+شرح ليا علاش خاصنا نديرو validation للبيانات قبل ما ندربو موديل
+ديال machine learning.
+
+
+[Calibration sample 6-45]
+Large language models are neural networks trained to predict tokens.
+During training, the model processes sequences of text and adjusts its
+parameters so that its predictions become closer to the expected next
+tokens. Modern transformer architectures use attention mechanisms to
+connect information across different positions in a sequence. During
+inference, the trained parameters remain fixed while the model generates
+new tokens step by step.
+
+
+[Calibration sample 6-46]
+Quantization reduces the numerical precision used to represent model
+weights. A model stored in BF16 may require substantially more memory
+than the same model represented with four-bit or two-bit quantization.
+Lower-bit quantization can make large models easier to deploy, although
+aggressive quantization can reduce accuracy and may require techniques
+such as importance matrices to preserve the most important information.
+
+
+[Calibration sample 6-47]
+Software engineering requires careful handling of errors, interfaces,
+dependencies, testing, and maintainability. A program that works for one
+input is not necessarily robust. Good implementations consider edge
+cases, validate assumptions, use clear abstractions, and provide useful
+error messages when something goes wrong.
+
+
+[Calibration sample 6-48]
+When evaluating an AI model, it is useful to measure different
+capabilities separately. Coding benchmarks can test program synthesis,
+mathematical datasets can test symbolic and numerical reasoning, and
+language evaluations can measure instruction following and multilingual
+performance. A single benchmark score does not completely describe the
+capabilities of a model.
+
+
+[Calibration sample 7-0]
+Write a Python function that determines whether a string is a palindrome.
+The function should ignore spaces, punctuation, and capitalization.
+Explain the algorithm and its time complexity.
+
+
+[Calibration sample 7-1]
+Implement binary search in Python for a sorted list of integers.
+Return the index of the target or -1 when it is not present.
+Explain why binary search runs in logarithmic time.
+
+
+[Calibration sample 7-2]
+Write a Python function that counts the frequency of every word
+in a text. Normalize capitalization and ignore punctuation.
+Return the words ordered from most frequent to least frequent.
+
+
+[Calibration sample 7-3]
+Given a list of integers, find the maximum subarray sum using
+Kadane's algorithm. Explain why the algorithm works and give its
+time and space complexity.
+
+
+[Calibration sample 7-4]
+Implement a stack using two queues. Explain each operation and
+the tradeoff between the possible implementations.
+
+
+[Calibration sample 7-5]
+Write a function that checks whether two strings are anagrams.
+The solution should run in O(n) time using a frequency dictionary.
+
+
+[Calibration sample 7-6]
+Explain the difference between a process and a thread in an
+operating system. Give an example where multithreading is useful.
+
+
+[Calibration sample 7-7]
+Write a recursive function to calculate the factorial of an integer.
+Handle zero correctly and explain the base case.
+
+
+[Calibration sample 7-8]
+Explain what an API is and show a small Python example of sending
+a JSON request to an HTTP endpoint.
+
+
+[Calibration sample 7-9]
+Design a simple REST API for a chat application. Describe endpoints
+for creating conversations, sending messages, and retrieving history.
+
+
+[Calibration sample 7-10]
+Solve the equation 3x + 7 = 31. Show every algebraic step.
+
+
+[Calibration sample 7-11]
+A rectangle has a length of 12 centimeters and a width of 7 centimeters.
+Calculate its area and perimeter.
+
+
+[Calibration sample 7-12]
+If a train travels 240 kilometers in 3 hours, calculate its average
+speed in kilometers per hour.
+
+
+[Calibration sample 7-13]
+A number is multiplied by 4 and then increased by 9 to produce 37.
+Find the original number.
+
+
+[Calibration sample 7-14]
+There are 5 red balls and 3 blue balls in a bag. What is the probability
+of randomly selecting a blue ball? Explain your reasoning.
+
+
+[Calibration sample 7-15]
+A shop gives a 20 percent discount on a product priced at 250 dirhams.
+Calculate the discount and the final price.
+
+
+[Calibration sample 7-16]
+If x + y = 18 and x - y = 6, solve for x and y using a system of equations.
+
+
+[Calibration sample 7-17]
+Explain why the square root of 2 is irrational using a proof by contradiction.
+
+
+[Calibration sample 7-18]
+A sequence begins 2, 4, 8, 16, 32. Describe the pattern and calculate
+the next three terms.
+
+
+[Calibration sample 7-19]
+A triangle has a base of 10 centimeters and a height of 8 centimeters.
+Calculate its area and explain the formula.
+
+
+[Calibration sample 7-20]
+A farmer has chickens and cows. There are 20 animals and 56 legs.
+Determine how many chickens and cows there are. Explain the reasoning.
+
+
+[Calibration sample 7-21]
+Three boxes are labeled apples, oranges, and mixed. Every label is wrong.
+You may take one fruit from one box. Explain how to correctly relabel
+all three boxes.
+
+
+[Calibration sample 7-22]
+A clock shows 3:15. What is the smaller angle between the hour hand
+and minute hand? Explain carefully.
+
+
+[Calibration sample 7-23]
+You have eight identical-looking balls and one is heavier.
+Using a balance scale only twice, explain how to find the heavier ball.
+
+
+[Calibration sample 7-24]
+A sequence contains 1, 1, 2, 3, 5, 8, 13. Identify the rule and
+calculate the next five values.
+
+
+[Calibration sample 7-25]
+Explain artificial intelligence to a ten-year-old using a simple
+real-world analogy. Avoid technical jargon where possible.
+
+
+[Calibration sample 7-26]
+Explain the difference between machine learning and traditional
+programming using a clear example.
+
+
+[Calibration sample 7-27]
+What is a neural network? Explain neurons, weights, training,
+and inference in simple but technically accurate language.
+
+
+[Calibration sample 7-28]
+Explain why training data quality can matter as much as training
+data quantity when building a language model.
+
+
+[Calibration sample 7-29]
+Describe the difference between an open-source model and an
+open-weight model. Explain why the distinction matters.
+
+
+[Calibration sample 7-30]
+اشرح مفهوم الذكاء الاصطناعي بطريقة بسيطة لطفل صغير، مع إعطاء أمثلة
+من الحياة اليومية وكيف يمكن للحاسوب أن يتعلم من البيانات.
+
+
+[Calibration sample 7-31]
+اشرح الفرق بين البرمجة التقليدية وتعلم الآلة، وقدم مثالا بسيطا
+يوضح كيف يتعلم النموذج من البيانات.
+
+
+[Calibration sample 7-32]
+حل المعادلة 3س + 7 = 31 واشرح جميع الخطوات بطريقة واضحة.
+
+
+[Calibration sample 7-33]
+ما هو النموذج اللغوي الكبير؟ اشرح كيف يتم تدريبه وكيف يستطيع
+توليد النصوص بعد التدريب.
+
+
+[Calibration sample 7-34]
+اشرح أهمية البيانات الجيدة عند تدريب نماذج الذكاء الاصطناعي
+ولماذا لا تعني كمية البيانات الكبيرة دائما جودة أفضل.
+
+
+[Calibration sample 7-35]
+شرح ليا شنو هو الذكاء الاصطناعي بطريقة بسيطة وبأمثلة من الحياة
+اليومية ديالنا فالمغرب.
+
+
+[Calibration sample 7-36]
+شنو الفرق بين البرمجة العادية وتعلم الآلة؟ شرحها ليا بطريقة سهلة
+وبمثال واضح.
+
+
+[Calibration sample 7-37]
+إلى كان عندنا 3x + 7 = 31، شحال كتساوي x؟ وشرح ليا المراحل وحدة بوحدة.
+
+
+[Calibration sample 7-38]
+شرح ليا شنو هو الموديل ديال اللغة الكبير وكيفاش كيتعلم من الداتا.
+
+
+[Calibration sample 7-39]
+علاش الداتا النظيفة والمزيانة مهمة فالتدريب ديال موديلات الذكاء
+الاصطناعي؟ وعلاش ماشي ديما الكمية الكبيرة هي الأفضل؟
+
+
+[Calibration sample 7-40]
+بغيت نبني تطبيق صغير بالبايثون كيجاوب على أسئلة المستخدمين.
+شنو هي المكونات اللي خاصني وكيفاش نقدر ننظم المشروع؟
+
+
+[Calibration sample 7-41]
+شرح ليا شنو هو API وكيفاش نقدر نستعملو فمشروع ديالي بالبايثون.
+
+
+[Calibration sample 7-42]
+عندي لائحة ديال الأرقام وبغيت نلقى أكبر مجموع ديال عناصر متتابعة
+فيها. شرح ليا كيفاش نخدم بخوارزمية Kadane.
+
+
+[Calibration sample 7-43]
+شنو الفرق بين CPU و GPU؟ وعلاش الـGPU مهم بزاف فالتدريب ديال
+موديلات الذكاء الاصطناعي؟
+
+
+[Calibration sample 7-44]
+شرح ليا علاش خاصنا نديرو validation للبيانات قبل ما ندربو موديل
+ديال machine learning.
+
+
+[Calibration sample 7-45]
+Large language models are neural networks trained to predict tokens.
+During training, the model processes sequences of text and adjusts its
+parameters so that its predictions become closer to the expected next
+tokens. Modern transformer architectures use attention mechanisms to
+connect information across different positions in a sequence. During
+inference, the trained parameters remain fixed while the model generates
+new tokens step by step.
+
+
+[Calibration sample 7-46]
+Quantization reduces the numerical precision used to represent model
+weights. A model stored in BF16 may require substantially more memory
+than the same model represented with four-bit or two-bit quantization.
+Lower-bit quantization can make large models easier to deploy, although
+aggressive quantization can reduce accuracy and may require techniques
+such as importance matrices to preserve the most important information.
+
+
+[Calibration sample 7-47]
+Software engineering requires careful handling of errors, interfaces,
+dependencies, testing, and maintainability. A program that works for one
+input is not necessarily robust. Good implementations consider edge
+cases, validate assumptions, use clear abstractions, and provide useful
+error messages when something goes wrong.
+
+
+[Calibration sample 7-48]
+When evaluating an AI model, it is useful to measure different
+capabilities separately. Coding benchmarks can test program synthesis,
+mathematical datasets can test symbolic and numerical reasoning, and
+language evaluations can measure instruction following and multilingual
+performance. A single benchmark score does not completely describe the
+capabilities of a model.