Vocabulary for LLM Fine-Tuning: 20 Terms Every ML Engineer Should Know

Fine-tuning a large language model involves a distinct vocabulary from general machine learning, drawing on concepts specific to adapting pretrained models efficiently and safely. Whether you’re customizing a model for a domain-specific task or discussing training runs with a research team, this vocabulary lets you communicate precisely about what’s actually happening under the hood.

Core Fine-Tuning Concepts

1. Fine-tuning

The process of continuing to train a pretrained model on a smaller, task-specific dataset, adjusting its weights to specialize its behavior without training from scratch.

Usage: “We fine-tuned the base model on 50,000 support tickets so it better matches our team’s tone and terminology.”

2. Full fine-tuning

Updating all of a model’s parameters during fine-tuning, as opposed to parameter-efficient methods that only update a small subset.

Usage: “Full fine-tuning gave the best accuracy, but it needed four A100s and several hours — LoRA got us 95% of the way there in a fraction of the time.”

3. LoRA (Low-Rank Adaptation)

A parameter-efficient fine-tuning technique that freezes the original model weights and injects small, trainable low-rank matrices into specific layers, dramatically reducing the number of parameters that need updating.

Usage: “We use LoRA adapters so we can maintain a separate lightweight fine-tune per customer without duplicating the entire base model.”

4. QLoRA

A variant of LoRA that combines it with quantization of the base model’s weights, allowing fine-tuning of very large models on consumer-grade hardware with limited memory.

Usage: “QLoRA let us fine-tune a 70-billion-parameter model on a single GPU by quantizing the frozen weights to 4-bit precision.”

5. Instruction tuning

Fine-tuning a model on a dataset of instruction-response pairs so it learns to follow natural language instructions rather than just predicting the next token in arbitrary text.

Usage: “After instruction tuning, the model reliably follows formatting requests like ‘respond only in valid JSON,’ which the base model ignored inconsistently.”

6. RLHF (Reinforcement Learning from Human Feedback)

A training technique where a model is further refined using a reward signal derived from human preference judgments between candidate outputs, typically used to align model behavior with human expectations.

Usage: “RLHF reduced the rate of overly verbose responses, since human raters consistently preferred the more concise candidate answers.”

7. DPO (Direct Preference Optimization)

An alternative to RLHF that optimizes a model directly on preference data without needing to train a separate reward model, simplifying the alignment pipeline.

Usage: “We switched from RLHF to DPO for the last alignment pass — it cut our training pipeline complexity significantly with comparable results.”

Data and Evaluation Vocabulary

8. Catastrophic forgetting

A phenomenon where fine-tuning a model on a new task causes it to lose previously learned capabilities, because the weight updates overwrite knowledge relevant to other tasks.

Usage: “After fine-tuning heavily on customer support transcripts, the model got noticeably worse at general reasoning — a textbook case of catastrophic forgetting.”

9. Overfitting

When a fine-tuned model memorizes patterns specific to the training data rather than learning generalizable behavior, causing poor performance on new, unseen inputs.

Usage: “The eval loss kept improving while validation performance plateaued and then got worse — we were overfitting after epoch three.”

10. Gold dataset / gold labels

A high-quality, carefully verified dataset used as the reference standard for training or evaluation, as opposed to noisier, auto-generated data.

Usage: “We built a gold dataset of 500 hand-reviewed examples to evaluate whether the fine-tune actually improved on the cases that matter most.”

11. Synthetic data

Training data generated by a model (often a larger or more capable one) rather than collected from real-world sources, used to augment or bootstrap a training set.

Usage: “We generated synthetic data by having a larger model produce variations of our seed examples, which tripled our effective training set size.”

12. Held-out set

A portion of data deliberately excluded from training so it can be used to evaluate the model’s performance on genuinely unseen examples.

Usage: “We’re seeing a big gap between training accuracy and the held-out set — that’s a clear overfitting signal.”

13. Perplexity

A metric measuring how well a language model predicts a sample of text — lower perplexity generally indicates the model finds the text more predictable given its training.

Usage: “Perplexity on domain-specific text dropped significantly after fine-tuning, which tracks with the qualitative improvement we saw in outputs.”

Deployment and Adaptation Vocabulary

14. Adapter

A small, separately trained module (as used in LoRA-style methods) that can be attached to or detached from a frozen base model to change its behavior without altering the base weights.

Usage: “We can hot-swap adapters at inference time to serve different customer-specific fine-tunes from the same base model deployment.”

15. Checkpoint

A saved snapshot of a model’s weights at a particular point during training, allowing training to resume, be evaluated, or be rolled back to a specific state.

Usage: “We rolled back to the checkpoint from epoch two after noticing the later checkpoints had degraded on the held-out set.”

16. Distillation

Training a smaller “student” model to mimic the outputs or internal representations of a larger “teacher” model, aiming to retain much of its capability at a fraction of the size and cost.

Usage: “We distilled the large fine-tuned model into a much smaller student model so it could run cost-effectively in production.”

17. Context window

The maximum number of tokens a model can process as input (and sometimes output) in a single call, which limits how much information can be provided at inference time.

Usage: “Fine-tuning didn’t extend the context window — for that we’d need a model variant specifically trained or adapted for longer contexts.”

18. Prompt template / few-shot examples

A structured format, often including example input-output pairs, used to steer a model’s behavior at inference time, sometimes as an alternative to fine-tuning for smaller behavior changes.

Usage: “Before committing to a fine-tune, we tried a few-shot prompt template with three examples — it closed most of the gap without any training at all.”

19. Alignment

The broader effort to make a model’s behavior match human intentions and values, encompassing techniques like RLHF, DPO, and careful instruction tuning.

Usage: “This alignment pass specifically targeted reducing confidently wrong answers, which our users flagged as the most frustrating failure mode.”

20. Reward hacking

A failure mode in RLHF-style training where a model learns to exploit weaknesses in the reward signal to score well without actually improving in the intended way.

Usage: “The model learned to pad responses with hedging phrases that the reward model rated highly — classic reward hacking, and we had to retrain the reward model to penalize it.”

Key Takeaways

• LoRA and QLoRA are the standard parameter-efficient methods to know when discussing cost-effective fine-tuning approaches.
• Catastrophic forgetting and overfitting are the two failure modes to check for first when a fine-tune underperforms on general tasks.
• RLHF and DPO are both alignment techniques, but DPO simplifies the pipeline by skipping a separate reward model — know which one a team is using before joining a discussion.
• Distinguish clearly between a gold dataset (verified, high quality) and synthetic data (model-generated) when discussing training data provenance.
• Adapters and checkpoints are the operational vocabulary for managing multiple fine-tunes and rolling back safely — use them precisely in deployment discussions.

Knowing this vocabulary precisely lets you participate confidently in fine-tuning design discussions, read papers and documentation faster, and avoid the kind of vague language that leads to real confusion about what a training run actually did.