A generative model can write something useful and at the same time introduce a false information. Learn how to separate how to generate, what receives context and how to check your response.
Objectives: explain tokens, parameters and care To compare transformed architecture; To design verifiable prompts and choose between prombing, RAG and fine-tuning.
Key concepts
A Large Language Model (LLM) models language patterns. In autoregressive generation, it takes context, gets scores for the potential tokens, select one and add it to the context to continue. It ends with an output condition, as an end token or a length limit.
A token It can be a word, part of a word or sign. tokenization depends on model and language. To count words doesn't mean to count tokens. The vocabulary does not disappear and is reconstructed in every sentence: the context modifies the probabilities of its elements.
The parameters they are learned values, as pesos. The tokens are input or output units, and are not parameters. The context window limits what an invocation can consider. The model size, numerical accuracy and context length have an impact on resources. more parameters do not best guarantee a better result in your task.
Reforms and embeddings
The self-attention allows to link positions from a sequence and ponder relevant information. The position representation brings the order. In "Ana put the equipment at the table because it was damaged," the context helps to link "damaged" with the equipment. That capacity did not provide a guarantee of proper understanding.
| Variant | Common paper | Example |
|---|---|---|
| Encoder | Produce text representations | Classification or embeddings |
| Decoder | Generate tokens conditioned by context | Writing or talking |
| Incoder- decoder | Transform one sequence into another | Translation |
The transformers allow some of the processing to be paralleled during training. That doesn't mean an autoregressive generation knows its future tokens. Due care prevents them from being used. A embedding represents content as a vector to compare learned relationships; it is neither a Spanish summary nor an encrypted copy of the document. Documents and queries must use a compatible embedding space.
Propting, RAG and fine-tuning
A clear prompt specifies task, context, restrictions and expected format. Zero-shot does not give examples. few-shot incorporates some. Learning in context conditions the response without changing weights. To demand a brief and verifiable explanation makes it easier to revise an outcome. does not give guaranteed access to internal reasoning or prove truthfulness.
Training tuning Training with instructions and responses to improve monitoring of tasks. RLHF uses human feedback to align conduct; doesn't amount to editing a prompt at every call.
RAG regains relevant fragments, incorporates them into the context and generates a response supported by them. It's useful for private policies or information that changes. Fine-tuning adapt a pre-trained model with examples, for example for a dedicated task or style. It can set all weights or a part of the mandateable parameters. It's not an automatic substitute for viewing current data.
It starts with a simple appraisal and prompt. It improves recovery if evidence lacks and considers fine-tuning if a behavioral problem remains that the examples and instructions do not solve. They can be combined. No obligation to use them all.
Step-by-step example
One policy says: "The preventive review is included during 90 days after delivery." The user asks about a dedicated computer 35 days ago.
- Reforms the current version and checks that the user can access it.
- It contributes the fragment and data from the equipment to the model, separate from the instructions.
- He called for a brief response with reference to politics and without creating conditions.
- Check the arithmetic:
90 − 35 = 55within the period, if there are no other conditions. - If date or conflicting versions are missing, she requests clarification or recognizes insufficient evidence.
A lower temperature usually reduces its variation. doesn't make a false fact true. A quote should also target a source that really supports the statement.
Common errors
- To merge an embedding with the textual response of the chat model.
- To say that few-shot re-trains weights or that fine-tuning automatically updates policies.
- To assume that all transformers have an obligation to encoder and decoder.
- Treat RAG as a guarantee of zero hallucinations: Recovery or generation may fail.
- Ensure that more context always helps, even if you introduce noise or bad instructions.
Retirement exercise
An application should respond with an inventory that changes every hour and use a short format. What would you prove before training a new model?
Solution
Query's authorized inventory using a tool and contributes that result to the context. use instructions and examples for the format. If information's in document, evaluate RAG. Fine-tuning can help with persistent behavior but doesn't get the model to know about this hour's stocks by itself. Check both accuracy data and format.
Sources and review
Check Generative AI and LLM Foundations at MyLearning. For architecture's origin, consult Attention Is All You Need. He relates these ideas to OCI Generative AI.