Audio, video, and LLM datasets

Sample-based projects for audio, video, LLM fine-tune curation, and RAG evaluation sets.

Modalities

Modality	Status	Project types
`audio`	`available`	`audio_classification`, `transcription`
`video`	`available`	`video_classification`, `action_recognition`
`llm`	`available`	`instruction_tuning`, `preference`, `rag_eval`

Audio

{"uri": "file:///data/clip.wav", "label": "speech"}
{"uri": "s3://bucket/a.wav", "transcript": "hello world"}

Video

{"uri": "file:///data/clip.mp4", "label": "walk", "fps": 30}

Instruction tuning

{"instruction": "Summarize", "input": "…", "output": "…"}

Preference

{"prompt": "Write a haiku", "chosen": "…", "rejected": "…"}

RAG eval

{"query": "…", "contexts": ["…"], "answer": "…", "relevant": true}

RAG eval is not a fine-tune task. Workflow: import → curate → clean → package / export JSONL for RAGAS, DeepEval, or a custom harness.

POST /api/projects/{id}/import/jsonl — any sample modality
POST /api/projects/{id}/export with format: jsonl
PATCH /api/samples/{id} — transcript, chosen/rejected, instruction/RAG fields, or class_id

Canonical export includes audio/video/llm samples.

Create project → choose Audio / Video / LLM types → Import JSONL → Annotate.

Media playback uses the URI when the browser can reach it (local file:// may be blocked; store http(s) or serve files separately).

Full video timeline editors, ASR workers, preference ranking UIs beyond chosen/rejected, GPU LLM fine-tuning.