---
license: cc0-1.0
task_categories:
  - other
tags:
  - video-encoding
  - vmaf
  - crf
  - x264
  - x265
  - vp9
  - av1
  - lightgbm
  - video-quality
size_categories:
  - 1M<n<10M
---

# CRF <-> VMAF training data

Training datasets behind [Alice-1](https://github.com/jakubkrapiec/alice-1) - a LightGBM model that predicts the CRF value needed to hit a target VMAF score for a given video segment, codec and resolution.

## Contents

| File | Rows | Description |
|---|---|---|
| `training_table.parquet` | 1,762,232 | **Main table.** Feature aggregates -> CRF for a target VMAF, per codec/resolution. |
| `probe_data.parquet` | 52,316 | **Probe-encode data.** Two 2s probe encodes per (segment, codec, resolution) cell with measured VMAF + bitrate. |

## Schema: `training_table`

One row = (video segment * codec * target resolution * target VMAF) -> the CRF that achieves that VMAF.

| Column | Type | Description |
|---|---|---|
| `source_key` | string | Source video identifier (`dataset/filename`) |
| `dataset` | string | Corpus subset (`main`, `hc`, `ext`, …) |
| `split` | string | Train/validation/test split |
| `segment_start` | float | Segment start time (s) |
| `codec` | string | `x264`, `x265`, `vp9`, `av1` |
| `preset` | string | Encoder preset (mostly `veryfast`; null for some rows) |
| `source_width` / `source_height` | int | Source resolution |
| `target_width` / `target_height` | int | Encoded resolution (720/1080/1440/2160) |
| `fps` | float | Frames per second |
| `segment_duration` | float | Segment length (s, ~10) |
| `vmaf_model` | string | VMAF model version |
| `fit_method` | string | How CRF was extracted from the encode ladder (`logistic` fit) |
| `si_mean` / `si_std` | float | Spatial Information (ITU-T P.910) mean/std over segment |
| `ti_mean` / `ti_std` | float | Temporal Information mean/std |
| `vmafmotion` | float | VMAF motion feature |
| `target_vmaf` | int | Target VMAF score (60–95) |
| `crf` | float | **Label:** CRF achieving `target_vmaf` |

## Schema: `probe_data`

One row = (segment * codec * resolution) with two cheap probe encodes.

| Column | Description |
|---|---|
| `probe_crf1`, `probe_vmaf1`, `probe_bitrate1_kbps` | First probe encode (2s): CRF, measured VMAF, bitrate |
| `probe_crf2`, `probe_vmaf2`, `probe_bitrate2_kbps` | Second probe encode (different CRF) |
| `probe_sec` | Probe duration (2.0s) |
| `probe_vmaf` | = `probe_vmaf1` (primary probe feature) |
| `probe_slope` | VMAF slope between the two probe points |
| `probe_log_br`, `probe_log_br2` | log10 of probe bitrates |

## Sources

~40,861 unique (segment, codec, resolution) cells from ~9,000 source clips: Blender Open Movies, Internet Archive public-domain films, NASA SVS, Wikimedia Commons, Pexels/Pixabay/Coverr/Mixkit stock, Xiph.org test media, YouTube CC. All sources are openly licensed.

## Citation

If you use this dataset, please link the repo.

## License

CC0-1.0 (public domain dedication). Feature aggregates and measurements computed by the authors from openly licensed sources; no video content is included.