AI and Dataset Glossary
This glossary defines the AI and data terms used by FishFarmHub internal dataset listings. Each definition explains the general concept and how the term appears in the current catalog.
Core AI and data concepts
Artificial intelligence (AI)
Artificial intelligence (AI) is the field of building computer systems that perform tasks such as language understanding, reasoning, classification, and generation.
FishFarmHub uses AI as an umbrella term for datasets prepared for language-model training or evaluation.
Related fields or listings: Related listing: Grouper Aquaculture AI Training Bundle.
Large language model (LLM)
A large language model (LLM) is a machine-learning model trained on text collections to process and generate language.
FishFarmHub datasets provide structured examples for training or evaluating language-model behavior in grouper aquaculture.
Related fields or listings: Related dataset types: QA, CoT, SFT, RAG, and DPO.
Question-answer pair (QA)
A question-answer (QA) pair is a record that connects one question with its corresponding answer.
FishFarmHub QA datasets store the question and answer as separate CSV fields; some variants also include reasoning or context.
Related fields or listings: Related fields: question and answer.
Comma-separated values (CSV)
A comma-separated values (CSV) file is a plain-text table where each row is a record and each column is a named field.
FishFarmHub delivers current internal datasets as CSV files or ZIP packages containing CSV files.
Related fields or listings: Related concepts: row, column, record, and field.
Chain-of-thought (CoT)
Chain-of-thought (CoT) reasoning is written intermediate reasoning that connects a question to an answer.
The FishFarmHub CoT dataset stores question, reasoning, and answer as separate fields.
Related fields or listings: Related fields: question, reasoning, and answer.
Supervised fine-tuning (SFT)
Supervised fine-tuning (SFT) is a model-training method that uses labeled inputs paired with desired outputs.
The FishFarmHub SFT dataset pairs each instruction with a response and supporting evidence.
Related fields or listings: Related fields: instruction, response, evidence, type, difficulty, and quality_score.
Retrieval-augmented generation (RAG)
Retrieval-augmented generation (RAG) is a workflow that retrieves relevant context before a model generates an answer.
The FishFarmHub RAG product is evaluation data containing questions, answers, positive contexts, and hard-negative contexts; it is not a hosted RAG system.
Related fields or listings: Related fields: question, answer, positive_context, and hard_negative_context.
Direct preference optimization (DPO)
Direct preference optimization (DPO) is a preference-training method that learns from a prompt paired with a preferred response and a rejected response.
The FishFarmHub DPO dataset stores prompt, chosen, rejected, critique, and quality_score fields.
Related fields or listings: Related fields: prompt, chosen, rejected, critique, and quality_score.
Dataset fields and evaluation concepts
Instruction
An instruction is the task or question supplied as model input in a supervised fine-tuning record.
FishFarmHub stores the user request in the instruction field of each SFT sample.
Related fields or listings: Related dataset type: SFT.
Response
A response is the desired answer paired with an instruction in a supervised fine-tuning record.
FishFarmHub stores the target assistant answer in the response field of each SFT sample.
Related fields or listings: Related fields: instruction and evidence.
Evidence
Evidence is source text stored with a record to support its response.
FishFarmHub SFT records use the evidence field to preserve text associated with the target response.
Related fields or listings: Related fields: instruction and response.
Positive context
Positive context is retrieved text that directly supports the answer to a RAG question.
FishFarmHub RAG evaluation records store supporting text in the positive_context field.
Related fields or listings: Related fields: question, answer, and hard_negative_context.
Hard-negative context
Hard-negative context is topically similar text that should not support the correct answer.
FishFarmHub RAG evaluation records use hard-negative context to test whether retrieval or answering distinguishes relevant evidence from a plausible distractor.
Related fields or listings: Related fields: question, answer, and positive_context.
Prompt
A prompt is the input question or instruction shared by a DPO preference pair.
FishFarmHub DPO records compare chosen and rejected responses to the same prompt.
Related fields or listings: Related fields: chosen, rejected, and critique.
Chosen response
A chosen response is the response marked as preferred in a DPO record.
FishFarmHub stores the preferred answer in the chosen field.
Related fields or listings: Related fields: prompt, rejected, and critique.
Rejected response
A rejected response is the less suitable response paired with the chosen response in a DPO record.
FishFarmHub stores the non-preferred answer in the rejected field.
Related fields or listings: Related fields: prompt, chosen, and critique.
Critique
A critique is a short explanation of why the rejected response is weaker than the chosen response.
FishFarmHub DPO records store this explanation in the critique field.
Related fields or listings: Related fields: chosen and rejected.
Quality score
A quality score is the numeric value stored with a sample to represent its recorded quality rating.
Current FishFarmHub SFT, RAG, and DPO records use a quality_score value from 1 to 5; the field does not imply an undisclosed external standard.
Related fields or listings: Related field: quality_score.