The popular eight-step infographic is directionally useful, but real-world model development is less linear, more iterative, and more technically nuanced. This version keeps the clarity while correcting the oversimplifications.
Think of these as major stages in a production pipeline—not eight perfectly isolated boxes. In practice, data work, evaluation, safety, post-training and deployment form feedback loops.
Training begins with large collections of text, code and other permitted data sources.
Raw corpora are transformed into a training mixture with carefully chosen proportions.
Text is encoded into discrete token IDs that the neural network can process.
A Transformer learns statistical structure by predicting tokens from context across massive datasets.
Supervised examples shape the pretrained model into a more useful assistant.
Human or model-derived preferences can be used to favor better responses.
Models are tested for capability, robustness, reliability and unwanted behavior.
The trained model is served on production infrastructure and generates tokens in response to user input.
A deployed assistant is not simply “the pretrained model.” The experience can include post-trained weights, system-level instructions, safety layers, retrieval, tools, memory, orchestration and product infrastructure around the core neural network.
The source infographic is an effective introduction, but some captions imply a cleaner, more deterministic process than actually exists.
It correctly identifies the major concepts: data preparation, tokenization, next-token pretraining, instruction tuning, preference alignment, evaluation, safety and production inference.
Real model development involves repeated iteration. Evaluation can change the data mixture; safety findings can alter post-training; deployment telemetry can influence later model versions.
High-end curation includes provenance, deduplication, quality scoring, language balancing, domain mixing, contamination analysis and privacy safeguards—not merely removing spam or “bad” content.
They pursue similar behavioral goals but use different optimization strategies. Treating “RLHF / DPO” as one single method hides an important technical distinction.
This is separate from training. Your prompt is processed at inference time by an already-trained system.
Your message may be combined with system instructions, conversation history and—when enabled—relevant tool or retrieval context.
The model repeatedly computes a probability distribution over possible next tokens and selects tokens according to the serving configuration.
Output can be streamed, checked against product rules, formatted and sometimes combined with external tool results before you see it.
A modern LLM is trained to model sequences of tokens, then post-trained and wrapped in a larger product system so that those learned representations can be turned into useful, controlled responses at inference time.