# LibreHPS — Data provenance

LibreHPS-4B was trained exclusively on permissively-licensed
human-preference data for text-to-image and text-to-video generation.
This document summarises the licence audit performed for every
upstream dataset in the training blend and the per-image generator
audit applied to filter the data before training.

## Datasets used

All datasets in the LibreHPS training blend are distributed under one
of: **MIT**, **Apache-2.0**, **BSD-3-Clause**, or
**CDLA-Permissive-2.0**. None impose non-commercial, no-derivatives,
or share-alike restrictions on ML-model training derivatives.

| Dataset | Upstream licence |
|---|---|
| `MizzenAI/HPDv3` | MIT |
| `zai-org/ImageRewardDB` | Apache-2.0 |
| `zai-org/VisionRewardDB-Image` | Apache-2.0 |
| `zai-org/VisionRewardDB-Video` | Apache-2.0 |
| `Rapidata/text-2-image-Rich-Human-Feedback` | CDLA-Permissive-2.0 |
| `Rapidata/sora-video-generation-alignment-likert-scoring` | CDLA-Permissive-2.0 |
| `Rapidata/text-2-video-human-preferences-veo3` | CDLA-Permissive-2.0 |
| `TIGER-Lab/VideoFeedback` | MIT |
| `data-is-better-together/open-image-preferences-v1` | Apache-2.0 |
| `fudan-generative-ai/LiFT-HRA-20K` | Apache-2.0 |

CDLA-Permissive-2.0 §3.1 jointly with §5.4 exempts "Results"
(including ML-model artefacts) from any downstream restriction. This
is the basis on which Rapidata CDLA-2.0 datasets are included.

## Per-image generator audit

A subset of upstream datasets — notably the Rapidata model-vs-model
splits — embed images generated by closed-source services (OpenAI,
Midjourney, Imagen, etc.). Each retained training pair carries an
audit flag `license_audit_flags.upstream_generator_restriction`
recording whether the upstream generator's ToS prohibits
redistribution of its outputs. Rows flagged `true` were dropped before
training.

**Generators dropped:**

- **Midjourney** — ToS prohibits redistribution of generated images.
- **DALL·E 3** (via OpenAI API) — older OpenAI ToS prohibits
  redistribution; only retained where the upstream dataset itself
  asserts generator-ToS compatibility.

**Generators kept:**

- **Stable Diffusion family** (SD 1.x, SD 2.x, SDXL, SD 3.x) — open
  weights.
- **FLUX.1**, **FLUX.1.1**, **FLUX 2 pro** — API ToS permits
  redistribution.
- **Imagen 3 / 4 / 4 Ultra** — Google generative-AI ToS permits
  redistribution of outputs.
- **HunyuanImage 2.1**, **Seedream 3**, **Kolors**, **Lumina** — open
  weights.
- **Recraft v2 / v3**, **Ideogram V2** — API ToS permits
  redistribution.

## Excluded datasets

These were reviewed and explicitly excluded:

- **Pick-a-Pic v1 / v2** — no declared HF licence; the underlying
  Dreamlike Photoreal 2.0 generator carries an OpenRAIL-M
  non-commercial clause.
- **HPDv2 (standalone)** — research-use only.
- **Midjourney Discord scrapes** — ToS prohibits redistribution.
- **PKU-Alignment/SafeSora\*** — CC-BY-NC-4.0 (non-commercial).
- **CodeGoat24/VideoFeedback** — a derivative of
  `TIGER-Lab/VideoFeedback`; the original was ingested instead.

## What this means for users

LibreHPS-4B is safe to use commercially. Every training-data row was
either drawn from a permissively-licensed dataset, or filtered out by
the generator-redistribution audit above. The model weights are
released under Apache-2.0 (see [`LICENSE`](LICENSE)) with no
inherited non-commercial or no-derivatives clauses from the training
data.