yogkul2000/Video-FLAIR-data

Video-FLAIR Training Data

Training data for Video-FLAIR: Not Whether to Reason, But How (NeurIPS 2026).

Paper | Code | Project Website

This repository contains annotations only. No videos, images, or extracted frames are included. You need to obtain the media from the original source datasets listed below, under their own licenses and terms.

Contents

file	rows	description
sft/sft_train.jsonl	127,234	mode-conditioned SFT traces in ms-swift format
sft/sft_val.jsonl	500	held-out SFT traces
rl/rl_train.jsonl	9,000	RL prompts with gold answers and answer types
media_index.jsonl	127,734	maps each sample id to its source dataset, source file, and frame timestamps

SFT traces by reasoning mode:

mode	train	val
DIRECT	38,632	177
CONCISE	78,381	288
DEEP	10,221	35

In 12,736 of the training traces the system prompt is the ADAPTIVE prompt, so the model learns to choose the mode itself. The prompt_mode field records which system prompt was used and flair_mode records the mode of the target response.

RL prompts by source mode: CONCISE 4,170, DIRECT 2,644, DEEP 2,186. Answer types: mcq_letter, number, math, open, yesno, bbox, segment, temporal_span, stvg, track.

Formats

All media paths are relative to a media root of your choice, for example frames/DIRECT/img016185/00.jpg.

SFT row:

{
  "messages": [
    {"role": "system", "content": "You are answering a multimodal question in DIRECT mode. ..."},
    {"role": "user", "content": "<image>Please provide the bounding box coordinate of the region this sentence describes: bright orange cow. ..."},
    {"role": "assistant", "content": "<DIRECT>\n<think>\n... I locate this bright orange cow [54, 550, 326, 859].\n</think>\n<answer>\n{\"boxes\": [54, 550, 326, 859]}\n</answer>\n</DIRECT>"}
  ],
  "images": ["frames/DIRECT/img016185/00.jpg"],
  "sid": "img016185",
  "flair_mode": "DIRECT",
  "prompt_mode": "DIRECT"
}

Video rows use "videos": [["frames/<MODE>/<sid>/00.jpg", "frames/<MODE>/<sid>/01.jpg", ...]], a list of frame lists, and the user turn starts with <video>Frame timestamps (seconds): ....

RL row:

{
  "messages": [{"role": "user", "content": "<video>Frame timestamps (seconds): ...\nQuestion: ...\nOptions:\nA. ...\nB. ..."}],
  "videos": [["frames/DEEPX/<sid>/00.jpg", "..."]],
  "solution": "C",
  "answer_type": "mcq_letter",
  "sid": "<sid>",
  "source_mode": "DEEP"
}

Reasoning traces use grounded spans inside <think>: <obj>name</obj><box>[x1,y1,x2,y2]</box><t>seconds</t>.

Media index row:

{"sid": "vid000000", "frames": "frames/DIRECT/vid000000", "kind": "video", "dataset": "OneThinker-train-data", "media": ["Tracking/ElysiumTrack/videos/1052971823.mp4"], "frame_ts": [0.69, 2.07, 3.44, ...]}

Source Data and Licenses

The questions, answers, and media come from the datasets below. Before using this data you must accept the license and terms of use of each source dataset and comply with them. This repository does not grant any rights to the source media.

dataset	samples	link
OneThinker-train-data	125,233	https://huggingface.co/datasets/OneThinker/OneThinker-train-data
MovieChat-1K	1,306	https://huggingface.co/datasets/Enxin/MovieChat-1K_train
HD-EPIC	933	https://hd-epic.github.io/
EgoProceL	262	https://sid2697.github.io/egoprocel/

Most samples come from OneThinker-train-data. In media_index.jsonl, media is the path of the source file inside the OneThinker-train-data release (or the MovieChat-1K release), the video id for EgoProceL, and the question type and video id for HD-EPIC.

Rebuilding the Frames

1. Download the source datasets from the links above.
2. For every row of media_index.jsonl, write frames to <media_root>/<frames>/NN.jpg with NN = 00, 01, ...
   • Video: one JPEG per entry of frame_ts, taken at that time in seconds from the source video, with the long side resized to at most 768 px. For clips (clip present), frame_ts is in the time base of the full source video.
   • Image: the source images in the order listed in media, with the long side resized to at most 1600 px.
3. Point the jsonl paths at your media root:

import json

def localize(src, dst, root):
    with open(src) as fi, open(dst, "w") as fo:
        for line in fi:
            r = json.loads(line)
            if "images" in r:
                r["images"] = [f"{root}/{p}" for p in r["images"]]
            if "videos" in r:
                r["videos"] = [[f"{root}/{p}" for p in v] for v in r["videos"]]
            fo.write(json.dumps(r, ensure_ascii=False) + "\n")

Training scripts are in the code repository under ms-swift/examples/flair.

Citation

@inproceedings{kulkarni2026videoflair,
  title={Video-FLAIR: Not Whether to Reason, But How},
  author={Kulkarni, Yogesh and Fazli, Pooyan},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2026}
}