These LLM training and adaptation techniques pages decide whether a method is the first step for a specialist task. They are not Hugging Face tutorials and they will not rank for the method name alone. If the brief is only an acronym, start with the “when this is wrong” page. The job of this index is to pick a first method — or to say that no method is first — and then send you to the spoke that owns that decision.
By Shen Pandi, Towards AGI. Reviewed September 2026. This page answers: LLM training and adaptation techniques.
How to pick a first method
Start with the constraint, not the acronym. If you cannot name the labels, the first method is not a method. It is a readiness review. If labels exist and the job is stable behaviour, supervised fine-tuning is the usual first training conversation. If the artefact you will store is an adapter, LoRA is how that SFT run is usually done. If the frozen base does not fit, QLoRA is the memory sibling. Preference work is later. Distillation is later. A full update is later and needs a written reason. Write that order on the brief before anyone names a GPU week.
That order is an operations sequence. It is not a quality league. We will not publish an llm.build ranking of these methods on your book. We will tell you which spoke to open so you stop booking the fashionable one first.
The spokes this index points at
Supervised fine-tuning owns labelled pairs and the moments you must not train yet. LoRA owns adapter handover and multi-adapter operations. QLoRA owns the memory method and the ways people use 4-bit to look thorough. Preference alignment owns pair cost after an SFT baseline. Distillation owns teacher-first order and both licences. Full fine-tune versus PEFT owns the written reason for touching every weight.
Continued pretraining, synthetic data and quantisation for serving are later pages. If your brief named those, this index will not pretend they already have a spoke. Fine-tuning versus RAG lives on the commercial LLM fine-tuning page. We did not ship a second comparison URL for that query.
No labels, or labels disagree: readiness, not a trainer.
Labels exist, job is behaviour: SFT, usually with LoRA.
Base does not fit: QLoRA after the task exists.
Two already-correct drafts, written bar still fails: preference later.
Teacher already meets the bar, you must shrink: distillation later.
LoRA failed a named metric: full update only with that reason written.
First is often “none”
The most useful first method is no training run. Uncleared files, two specialists marking the same input differently, facts that change every week, nobody owning the refuse path — those are readiness and retrieval problems. Booking SFT, LoRA, QLoRA, DPO or a student will copy the problem at GPU prices. This index exists so a sponsor who already booked an acronym has a page that says stop.
If you are unsure whether any training is first, read supervised fine-tuning. That spoke is the gate. The other spokes assume you can pass it, or they will send you back.
What we will not do here
No copy-paste training scripts. No unpublished accuracy tables. No claim that a method moves inference or rewrites a foundation licence. Those sentences belong on the spokes in their own words. Repeating them here as a sermon would trip the overlap gate and would not help a buyer pick a URL.
This index also will not settle family choice. Llama versus Qwen is not a technique. Open weights versus a closed API is not a technique. Use the models catalogue and the compare index for those forks. Bring one constraint per URL.
Practice: a claims lead already booked QLoRA
The index is used to unbook the acronym
insurance
A claims lead opens this index because the steering pack already says QLoRA. They read the first-method order, not the acronym glossary. The honest sentences are: letters are not cleared, two handlers still fight about tone, there is no unused pack. The first URL is supervised fine-tuning (do not train yet), then data readiness. QLoRA is the wrong tab until a 70B-class base does not fit for ordinary LoRA on a box they operate.
They leave with two links and a refusal to keep QLoRA as the first engagement. That is a successful use of this index. It is not a training run. It is not a quality score. It is a buyer using the map to stop a booking.
Practice: a catalogue lead wants DPO because labs use it
The index is used to put preference after SFT
retail
A merchandising lead opens this index because a vendor said DPO is what serious teams do. They already have an allowed-value list for attributes. They do not yet have unused SKUs reserved. The first-method order says: unused pack, then SFT (usually LoRA), then preference only if two already-correct blurbs still fail a written copy bar.
They open supervised fine-tuning for the pack-and-pairs gate, and they bookmark preference alignment as a later spoke. They do not open distillation. They do not open a full update. The index did its job when the next meeting drops DPO as the first line.
How the spokes share a brief without sharing a sermon
Each spoke has a who-for, a limit box, insurance and retail examples, a “what changes in the run” section, decisions, and a handover or operating callout. This index will not reprint those. If you need the run mechanics, open the spoke. If you need the first-method order, stay here until you can name the constraint in one sentence.
Related links at the bottom are the spokes themselves, titled as their page titles. Use those titles. A buyer who says “we need the PEFT page” should land on full fine-tune versus LoRA, not on a blog that uses PEFT as a synonym for everything.
Memory, preference, shrink, every-weight — later forks
Memory is QLoRA after the task exists and the named base does not fit. Preference is DPO (or a neighbour) after an SFT baseline and pair hours you can actually buy. Shrink is distillation after a teacher already meets the bar and both licences allow traces. Every-weight is a full update after LoRA failed a named metric. Those four forks are how sponsors skip the gate. This index exists to put them back in order.
If two forks are in the same brief — “QLoRA then distill then DPO” — you have a programme, not a first method. Write the first method only. Programmes that start with three acronyms usually have no unused pack.
Decision prompts, not scores.
When
Then
You cannot name the labels or the unused pack
No training method. Open supervised fine-tuning and data readiness.
Labels exist; you need adapters on a base that fits
SFT with LoRA. Open those two spokes. Do not open DPO first.
The only new fact is “the base does not fit”
QLoRA after the task exists. Do not treat 4-bit as a quality tier.
Someone booked DPO, a student, or a full update first
Use this index to unbook it. Open the spoke that explains why it is later.
A first-method sentence you can actually write
Before you leave this index, write one sentence a specialist would recognise. Good: “We do not train until handlers share a letter standard and unused test documents exist.” Good: “Labels agree on attributes; first training is SFT with LoRA on a base that already fits.” Good: “Teacher already meets the extract bar; we are opening distillation only because the box cannot keep that teacher warm.” Bad: “We will QLoRA, DPO and distill in the same sprint.” Bad: “We will pick the method after we see which acronym the vendor likes.”
If you cannot write a good sentence, you are not ready for a spoke. You are ready for readiness or for a scoping call. This index is allowed to end in a stop. A stop is a first method.
How the spokes refuse each other’s jobs
SFT refuses a tone fight and a folder with no pairs. LoRA refuses a mystery merge and a licence disappearing act. QLoRA refuses 4-bit as readiness and 4-bit as a quality tier. Preference refuses a messy inbox and RLHF as the default desk path. Distillation refuses a public student filed as your run, and it refuses Llama traces into another family. Full versus PEFT refuses a full update that is only extra cost after LoRA already met the bar.
This index refuses the programme that books all six. Open one spoke. Finish that decision. Then, if the next constraint is real, open the next spoke. Buyers who collect all six titles into one week usually have no unused pack and no operator.
How a buyer should walk the related list
The related links at the bottom are the spokes, titled as their page titles. Read the title, not the acronym you already booked. “When supervised fine-tuning is the wrong first step” is the gate. “LoRA adapters for specialist models” is the artefact. “When QLoRA is the wrong first step” is the memory misuse. “DPO vs SFT for a specialist business task” is pair cost after a baseline. “Distilling an LLM into an SLM for a held-out task” is teacher-first. “Full fine-tune vs LoRA for specialist work” is the every-weight reason.
If you cannot match your constraint to one of those titles, you are not looking for a technique. You are looking for a family, a comparison, or a commercial page. Use the other indexes. This one will not invent a seventh spoke to make a vendor acronym feel at home.
Print the related list into a steering pack only if each title has a one-line verdict next to it: first, later, or not this brief. A pack that lists all six as “in scope” is the same programme this index refuses. Later is allowed. First is one.
If two people write two first-method sentences, you do not yet have a first method. You have two briefs. Split the meeting. The claims tone fight and the catalogue attribute job are not one SFT run, and they are not one QLoRA booking. This index is a map. It is not a blender.
What this index does not replace
It does not replace the commercial fine-tuning page, where RAG versus training still lives. It does not replace family pages. It does not replace the compare index. It does not score your unused pack. Buyers who treat this URL as a tutorial will be disappointed. Buyers who treat it as a first-method map will leave faster.
If you arrived from a search for one acronym, skip to that spoke. If you arrived from a search for “LLM training and adaptation techniques”, stay here long enough to write the first-method sentence, then leave. Lingering to collect every acronym is how this page gets misused as a course outline.
A first-method map that is used as a week-one course will produce six half-read spokes and no unused pack. That is worse than reading one spoke slowly. If a new joiner needs orientation, give them this page and the readiness review, then stop. Training vocabulary without a task is how acronyms get booked.
If the commercial page already answered “train versus retrieve”, do not open SFT to relitigate retrieval. Come here only when a training method is actually in scope. The index is a fork, not a second homepage. A homepage tour that opens every spoke is how first-method discipline dies in the first hour. One spoke, one verdict, then stop or book the next constraint.
Questions we hear first
Which technique should we start with?
Usually none until the task and unused-test plan exist. After that, SFT with LoRA is the common first training method. QLoRA when memory is the constraint. DPO later. Distillation after a teacher already meets the bar. A full update only after LoRA failed a named metric. If someone already booked three of those for the same week, unbook two.
Where is fine-tuning versus RAG?
On the commercial LLM fine-tuning page, not as a second comparison URL. We did not ship a commodity blog post for that query.
Is this a tutorial for the methods?
No. Each spoke is a decision page. Hugging Face and the papers are the tutorials. We will not paste scripts here.
Why is QLoRA a separate spoke from LoRA?
Because people book 4-bit as a quality tier or as a way to skip readiness. The QLoRA page owns those misuses. LoRA owns handover and multi-adapter ops.
When does preference work become first?
Almost never on a specialist extract or classification job. It becomes in-scope after an SFT baseline and a written bar that ranking can move.
When does distillation become first?
It does not. Teacher-first means the large model is already measured. A public student is not your run.
When is a full update the first training method?
Almost never. Default to LoRA. A full update needs a failed adapter metric written down.
Can a method create sovereignty?
No. Methods change artefacts. Inference location and operators live on the sovereignty checklist and the closed-API comparison. An adapter on a laptop is still not a rota.
Bring the constraint, not the acronym.
If someone already booked QLoRA or DPO, tell us why. We will say whether that method is the first step or a later one.