Large Language Models: A Plain-Language Overview for Engineers
Organizations: Department of Artificial Intelligence and Data Engineering, Ankara University, Ankara, Türkiye
Abstract
Large language models (LLMs) are neural networks trained on huge text collections to generate text by predicting what comes next. Engineers using ChatGPT and other generative artificial intelligence (AI) tools need a clear picture of what training changes and prompts add. The article starts with the Transformer, which connects words in context, and pre-training, where models learn to predict text. It then explains how fine-tuning adapts models to requests, with reinforcement learning from human feedback using human answer rankings. Next, retrieval-augmented generation (RAG) supplies documents, while chain-of-thought prompting encourages intermediate steps and agents use tools. FlashAttention lowers training and running costs by reducing data movement during attention; quantized low-rank adaptation (QLoRA) saves fine-tuning memory. Readers get a life cycle figure and tables comparing methods and their limits. A short checklist for judging LLM evaluation claims flags test material in training (data contamination) and model-based grading (LLM-as-a-judge). It also covers very long inputs (long context) and hallucination, meaning false or unsupported claims. Check factual accuracy separately from whether an answer sounds helpful.
Figures & tables
| Method | Supervision | Training work | Main limitation stated in the papers |
|---|---|---|---|
| Supervised instruction tuning [ 2 ] | Human-labeler supervision | Supervised fine-tuning | Labelers are not representative; data is almost entirely English |
| RLHF [ 2 ] | Human comparisons of model outputs | Fit preference-predicting reward model; optimize outputs using reinforcement learning | Narrow labeler population; benchmark regressions persist despite mixing pretraining updates |
| DPO [ 10 ] | Pairs of preferred and dispreferred responses | Direct preference updates; no separate reward model or reinforcement-learning loop | Larger-scale testing remains open; generalization evidence is preliminary |
| QLoRA [ 11 ] | OASST1 and subsampled FLAN datasets in chatbot experiments | Adapter fine-tuning across all linear transformer layers | Full fine-tuning parity unestablished at larger evaluated scales |
| Method | Stage | What it saves | Reported result | Caveat |
|---|---|---|---|---|
| Compute-optimal scaling (Chinchilla) [ 8 ] | Pre-training | Model size for a fixed compute budget | Same training compute as Gopher with a model four times smaller, so memory and inference cost drop | Only two large comparable runs; no tests at intermediate scales |
| FlashAttention [ 15 ] | Training and inference | Reads and writes to GPU main memory | GPT-2 medium attention: 40.3 GB of memory traffic cut to 4.4 GB; 41.7 ms cut to 7.3 ms | Each attention variant needs a hand-written GPU kernel; analysis covers one GPU |
| QLoRA [ 11 ] | Fine-tuning | GPU memory for weights and optimizer state | 65-billion-parameter model fine-tuned in under 48 GB; 4-bit with double quantization matched 16-bit on MMLU | Match with full 16-bit fine-tuning not shown at 33B and 65B |
| DPO [ 10 ] | Preference tuning | A separate reward model and the reinforcement-learning loop | Win rate about 61% vs 57% for PPO on summarization, judged by GPT-4 | Tested only up to 6B parameters |
| Check | Evidence to request | What remains unproven | Motivating finding |
|---|---|---|---|
| Test contamination [ 13 , 16 ] | Benchmark version, problem dates, training cutoff, and post-cutoff scores | Post-cutoff subsets can be noisy; text-overlap filters cannot exclude paraphrased tests | Some tested models scored sharply lower on problems released after their cutoffs |
| LLM-as-judge validity [ 7 , 10 , 11 ] | Answer-order swaps, length-bias tests, and calibration against human votes | Human agreement does not guarantee correctness; judge results depend on prompts | Repetitive verbosity fooled tested judges, with substantially different failure rates |
| Long-context use [ 14 ] | Accuracy across evidence positions and context lengths, not merely window size | Generality beyond the tested question-answering and synthetic retrieval tasks remains untested | Extended windows did not improve accuracy when both versions accommodated the input |
| Factual grounding [ 4 ] | Verify that retrieved passages support each answer claim; assess source quality | Retrieved sources may contain factual errors and bias | Mismatching index dates and question dates reduced accuracy in the world-leader test |
| Inference budget and reasoning [ 5 , 12 , 13 ] | Request token counts, sample counts, relevant tool calls, latency, and accuracy | Plausible reasoning traces do not guarantee correct reasoning | Competition-math evaluation reported thinking-token usage; simpler questions can trigger overthinking |
| Alignment and impact scope [ 2 , 6 ] | Evaluator population, task coverage, compute, training-data documentation, and environmental disclosures | Narrow truthfulness tests and incomplete risk audits cannot establish broad safety | InstructGPT used a nonrepresentative labeler population and remained incompletely aligned and unsafe |
LLM Training
Large language model training studies the data, objectives, optimizers, and systems used to build and adapt language models.
See all 344 papers