Curating Training Data for LLMs

Master Table

All training stages, dataset types, and their characteristics in one view.

Stage	Dataset Type	Turn Type	Format	Signal	Trains	Behaviors	Risks	Typical Style	Key Metadata
Pre-training	Raw text corpus	Document chunks	Books, web, code, docs	Next-token	Language, knowledge, syntax	Fluency, recall, summarization	Memorization, bias, stale data	Long-form prose	Source, language, quality score, dedup cluster
	Code corpus	Code files	Repos, notebooks, docs	Next-token	Programming patterns, APIs	Code completion, debugging	Insecure code, license issues	Structured code files	Language, repo, license, path
	Multimodal paired data	Cross-modal	Image/audio/video + text	Cross-modal training	Vision-language/audio-language	Grounded multimodal understanding	Alignment noise, OCR errors	Descriptive or grounded	Modality tags, alignment labels
	Structured-to-text	Document chunks	Tables/JSON as text	Next-token	Schema familiarity	Structured data comprehension	Fragile formatting habits	Serialized tabular data	Schema version, canonicalization
Continued Pre-Training (CPT)	Domain corpus	Document chunks	Legal, medical, finance	Next-token	Domain adaptation	Terminology, domain recall	Overfitting, domain bias	Technical or specialized	Domain tag, freshness, authority
Supervised Fine-Tuning (SFT)	Prompt-response	Single-turn	User prompt + answer	Supervised	Instruction following	Obedience, formatting	Over-verbosity	Assistant answer	Task type, gold answer, style rubric
	Dialogue tuning	Multi-turn	Full conversation	Supervised	Conversational coherence	Context tracking, follow-up	Drift, weak long-context	Conversational	Role labels, conversation state
	Task-specific	Single/structured	Input-output pairs	Supervised	Task accuracy	Classification, extraction, SQL	Narrow capability	Formal, constrained	Schema version, validator, label ontology
	Contrastive retrieval	Single-turn	Query + pos/neg docs	Contrastive loss	Retriever/reranker quality	Better search, RAG grounding	Hard negative selection critical	Search relevance	Hard negatives, doc metadata
Preference	Ranked answers	Single/multi-turn	Prompt + 2+ responses	Preference loss	Quality judgment	Helpfulness, concision	Annotator taste, bias	Comparative	Annotation policy, rater agreement
	Scalar scoring	Single/multi-turn	Prompt + score	Reward regression	Reward signal for RL	Quality estimation	Reward hacking	Evaluative	Rubric definitions, score ranges
	Constitutional critique/revision	Multi-step	Response + principle + revision	Supervised / preference	Self-critique, policy alignment	Principled self-improvement	Can become rigid	Editorial / self-correction	Principle set, critique rubric
Reinforcement Learning (RL)	Trajectories	Multi-step	State-action-reward	Policy optimization	Sequential decisions	Tool use, planning, browsing	Instability, exploitation	Interactive	Reward source, trajectory length
Safety/Tools	Tool-use training	Multi-turn	Actions + tool results	SFT + RL	External actions	API selection, formatting	Wrong tool, malformed calls	Agentic, precise	Tool schema, tool version, success flag
	Retrieval-grounded	Single/multi-turn	Query + context + answer	Supervised/preference	Grounded answering	Citations, less hallucination	Over-reliance on retrieval	Evidence-grounded	Document IDs, evidence spans, freshness
	Safety tuning	Single/multi-turn	Safe/unsafe + response	SFT + preference + RL	Boundary behavior	Refusals, de-escalation	Over-refusal	Careful, policy-aligned	Risk category, policy tags, severity