RLHF, from feedback to fine-tuning.
Reinforcement learning from human feedback, made simple. Collect, score, and export human feedback for your LLMs and AI agents with one API — and get a training-ready preference dataset in minutes.
What is RLHF?
RLHF — reinforcement learning from human feedback — is the technique behind aligned models like ChatGPT and Claude. Instead of training an LLM on raw text alone, you show humans model outputs, let them rate or rank which response is better, and use that preference data to train a reward model that steers the LLM toward more helpful, accurate, and safe answers.
The hard part isn't the theory — it's the data pipeline: capturing every interaction, attaching a human feedback score, keeping it clean, and exporting it in a training-ready format. That's exactly what the AINative RLHF API does, so you can go from collecting human feedback to fine-tuning data without building the plumbing yourself.
The RLHF loop, as an API
Collect human feedback, score it, export a dataset — three calls, no pipeline to build.
Log any prompt/response pair with a human feedback score (0.0–1.0) or a thumbs up/down. One POST captures it as structured preference data.
Filter interactions by feedback type, model, reward score, or tags. Batch-update ratings and add human notes to build a clean preference dataset.
Export a training-ready dataset in JSON, CSV, Parquet, or OpenAI fine-tuning format — ready to fine-tune, align, or run your own reward model.
Collect feedback in one call
# 1. Log a human-rated interaction (preference data)
curl https://api.ainative.studio/v1/public/$PROJECT_ID/database/rlhf/interaction \
-H "X-API-Key: $AINATIVE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"prompt": "Explain quantum computing simply.",
"response": "Quantum computing uses qubits that can be 0 and 1 at once...",
"feedback": 0.9,
"agent_id": "gpt-4",
"tags": ["science", "explanation"]
}'
# 2. Export a training-ready dataset (JSON, CSV, Parquet, or OpenAI)
curl https://api.ainative.studio/v1/public/$PROJECT_ID/database/rlhf/export \
-H "X-API-Key: $AINATIVE_API_KEY" \
-d '{"format": "openai", "reward_min": 0.8}'Export training-ready datasets
Filter by feedback, model, reward score, or tags — then export in the format your trainer wants.
JSON
CSV
Parquet
OpenAI
Feedback that feeds the whole platform
RLHF preference data lives alongside your database, your agent memory, and your analytics — one platform, no ETL between separate systems. Human feedback captured today becomes fine-tuning data, memory, and an analyzable dataset tomorrow.
The knowledge layer your agents provision themselves. RLHF preference data lives right next to your app data.
/api/v1/projects/{id}/database/*Semantic recall and a knowledge graph — feedback on agent runs feeds straight back into how your agents improve.
/api/v1/public/memory/v2/*Run free, unmetered SQL over your exported RLHF datasets to analyze reward distributions and label quality.
/database/lakehouse/queryFrequently asked questions
RLHF stands for reinforcement learning from human feedback — a technique for aligning LLMs with human preferences. Humans rate or rank model outputs, and that preference data trains a reward model that steers the LLM toward more helpful, accurate, safe responses. It is how models like ChatGPT and Claude were tuned to follow instructions.
It lets you collect human feedback on any AI interaction, score it, and export a training-ready dataset. POST a prompt/response pair with a feedback score (0.0–1.0) or thumbs up/down, and AINative stores it as structured preference data you can filter, review, and export as JSON, CSV, Parquet, or OpenAI fine-tuning format.
POST to /v1/public/{project_id}/database/rlhf/interaction with the prompt, the model response, and a feedback score. AINative captures it as preference data with metadata (model, tags, reward score, human notes). Every interaction is scoped to your own project.
JSON, CSV, Parquet, and OpenAI fine-tuning format. Export the full dataset or filter by feedback type, model, reward score, or tags — so you get exactly the human-feedback data you need to fine-tune or align your model.
Yes — you can start collecting human feedback for free. RLHF data collection is part of ZeroDB, the AINative knowledge layer for AI agents. Sign up, provision a project, and start logging interactions with no setup.