AI Business Process · Topic 1 of 2
AI and Large Model Business Processes
Learn the AI project lifecycle and the large model process: pre training, SFT, LoRA, RLHF alignment, quantization, distillation and deployment for H13-311.
- 5 min
- AI process: business understanding, data collection, data processing, training, evaluation, deployment, operation and iteration.
- Deployment can be cloud, edge or device; real time or batch.
- Large model process: pre training, supervised fine tuning, alignment, compression, deployment, application.
What it is
The AI business process is the lifecycle of an AI project: understand the business problem, prepare the data, develop and evaluate a model, deploy it and keep it healthy. Most of the effort goes into data, not into the model itself.
The large model business process follows the same idea but starts differently. Instead of training a small model for one task, you take a foundation model pretrained on huge data and adapt it to your business with fine tuning, alignment and compression.
How it works
Stages of the AI business process
| Stage | Key work |
|---|---|
| Business understanding | Define the goal, the task type, the metric and whether AI is the right tool |
| Data collection | Gather data from databases, logs, sensors or public datasets |
| Data processing | Clean (missing values, duplicates, outliers), label, augment and split the data |
| Model training | Select an algorithm, train it and tune hyperparameters |
| Model evaluation | Check metrics on the test set and confirm they meet the business goal |
| Deployment | Release the model as a service in the cloud, at the edge or on a device |
| Operation and iteration | Monitor accuracy, detect data drift and retrain with new data |
Data labeling is often the most expensive step, because supervised learning needs a correct label for every sample. Data drift means the live data slowly differs from the training data, so accuracy falls even though the model has not changed.
Huawei Cloud ModelArts is a one stop AI development platform that supports these stages, from data labeling to training and deployment.
Deployment options
| Option | Where inference runs | Strength |
|---|---|---|
| Cloud | Data center servers behind an API | Large models, easy updates |
| Edge | Gateways or servers near the data source | Low latency, less bandwidth |
| Device | Phones, cameras, vehicles | Works offline, keeps data private |
Real time services answer each request immediately. Batch services process many records on a schedule.
The large model lifecycle
| Stage | What happens | Data needed |
|---|---|---|
| Pre training | The model learns language and knowledge by predicting the next token, a self supervised task | Very large unlabeled corpora; very high compute |
| Supervised fine tuning (SFT) | The model learns to follow instructions from prompt and answer pairs | Smaller, high quality labeled examples |
| Alignment | The model is tuned toward helpful, honest and safe answers, for example with RLHF | Human preference rankings |
| Compression | The model is made smaller and faster for serving | No new task data, or a small calibration set |
| Deployment | The model runs as an inference service | Prompts from users |
Most businesses never pre train. They choose an existing base model and do only fine tuning, compression and deployment.
Fine tuning: full and parameter efficient
Full fine tuning updates every weight, which needs a lot of GPU or NPU memory. Parameter efficient fine tuning (PEFT) freezes the original weights and trains a small number of new ones.
LoRA (low rank adaptation) is the best known PEFT method. For a frozen weight matrix W of size d x k, it trains two small matrices B (d x r) and A (r x k) with a small rank r, and uses W + BA.
Alignment with RLHF
RLHF (reinforcement learning from human feedback) has three steps. Humans rank several answers to the same prompt, a reward model is trained on those rankings, and the language model is optimized with reinforcement learning (commonly PPO) to earn a high reward.
Compression
| Method | Idea |
|---|---|
| Quantization | Store weights with fewer bits, such as INT8 or INT4 instead of FP16 |
| Knowledge distillation | A small student model learns to copy a large teacher model |
| Pruning | Remove weights or structures that contribute little |
A worked example
A team adapts a model layer with a 4096 x 4096 weight matrix, which has 16,777,216 weights. LoRA with rank r = 8 trains B (4096 x 8) and A (8 x 4096), so 32,768 + 32,768 = 65,536 weights.
That is 65,536 / 16,777,216 = 0.39% of the original. The base weights stay frozen and can be shared by several LoRA adapters.
For serving, a 7 billion parameter model stored in FP16 (2 bytes per weight) needs about 14 GB just for the weights. INT8 quantization cuts this to about 7 GB, and INT4 to about 3.5 GB.
How the exam asks about this
- "Which stage of the AI business process usually takes the most effort?" → Data collection and processing, including labeling.
- "Which method freezes the base model and trains small low rank matrices?" → LoRA, a parameter efficient fine tuning method.
- "Which technique stores model weights with fewer bits to reduce memory?" → Quantization.
What to remember for the exam
- AI process: business understanding, data collection, data processing, training, evaluation, deployment, operation and iteration.
- Deployment can be cloud, edge or device; real time or batch.
- Large model process: pre training, supervised fine tuning, alignment, compression, deployment, application.
- Pre training is self supervised next token prediction on huge unlabeled data; SFT uses labeled instruction data.
- LoRA trains small low rank matrices; RLHF uses human rankings, a reward model and reinforcement learning.
Fine tuning changes the model's weights; prompt engineering and RAG do not. If a question says the knowledge must be updated daily without retraining, the answer is RAG, not fine tuning.
Common mistakes
Thinking every company pre trains its own large model. Pre training needs enormous data and compute. Most projects start from an existing foundation model and only fine tune or prompt it.
Treating deployment as the end of the project. Data drift slowly lowers accuracy. Monitoring and retraining are part of the lifecycle.
Confusing distillation with quantization. Quantization keeps the same model but stores its weights with fewer bits. Distillation trains a new, smaller model to imitate the large one.
Check yourself
What is data drift?
Data drift is when the live data a model sees gradually differs from its training data, so accuracy drops and the model must be retrained.
What kind of data does supervised fine tuning use?
High quality labeled pairs of instructions or prompts and the desired answers, far fewer than the pre training corpus.
Why is LoRA cheaper than full fine tuning?
It freezes the original weights and trains only two small low rank matrices per adapted layer, so far fewer parameters need gradients and optimizer memory.
Topic 1 of 2 in AI Business Process
Test yourself on AI Business Process
The full Huawei HCIA-AI H13-311 question bank has 600 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Huawei. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.