av-skills

AV-Skills Audio-visual instruction and temporally grounded reasoning data for Nemotron-Labs-Audio-Visual Flamingo AV-Skills supports joint understanding of video, speech, sound, music, and long-range temporal context in real-world videos. Dataset Summary AV-Skills is the audio-visual instruction and reasoning dataset for Nemotron-Labs-Audio-Visual Flamingo**, an open audio-visual language model for long and complex videos. It is designed for training models that jointly understand visual events, speech, environmental sounds, music, and temporal structure in real-world videos. This repository contains a normalized Hugging Face release combining: **AV-Skills**, a large-scale collection of captioning and QA conversations for joint audio-visual understanding. **AV-Think**, a temporally grounded reasoning dataset used for Temporal Audio-Visual Interleaved Chain-of-Thought (TAVIT). Nemotron-Labs-Audio-Visual Flamingo uses AV-Skills as part of a three-stage curriculum for open audio-visual LLM training: short-context SFT, long-context SFT, and CoT post-training. This repository packages the merged annotation release as 2,862,653 conversation rows**. Dataset Statistics | Category | Value | | --- | ---: | | Total conversation rows | 2,862,653 | | AV-Skills rows | 2,839,035 | | AV-Think rows | 23,618 | | Normalized JSONL shards | 11 | | Original task labels | 29 | | AVF maximum training video duration | 15 minutes | | AV-Think reasoning samples | roughly 24K | | AV-Think average reasoning-chain length | 635.7 words | The normalized data is stored under `data/`. Original raw JSON files from the source AV-Skills repository are also preserved at the repository root for traceability, but the Dataset Viewer configs point to the normalized JSONL files. Dataset Configs | Config | Split | Rows | Path | Description | | --- | --- | ---: | --- | --- | | `AV-Skills` | `train` | 2,839,035 | `data/AV-Skills/*.jsonl.gz` | Captioning and QA conversations from AV-Skills source shards. | | `AV