Bora

A curated, clustered video clip dataset built to train a from-scratch, small-parameter (~150-200M) video generation model as part of the Codeically "smallest video gen model" video series.

Contents

The repo has two layers:

Raw footage (batch_01/ … batch_NN/) — source videos as collected, each batch with its own manifest.csv:

column	description
filename	source video filename within the batch
source	pexels, pixabay, or nasa
category	one of 56 fine-grained domain tags (see below)
query	the search query used to find the clip
description	short human description (manually filled in for NASA clips; auto-filled elsewhere)
license	license as stated by the source platform
original_url	link to the original source page, for attribution/verification

Processed clips (clips/<cluster>/*.mp4 + clips_manifest.csv) — the actual training data. Each raw source video is cut into fixed-length windows, resized, and resampled:

• 2 seconds per clip, no overlap between windows
• ~160px wide (height auto-scaled), landing in a 128–192px band
• 10fps (20 frames per clip), landing in an 8–12fps band

clips_manifest.csv columns: clip_filename, cluster, category, source, license, description. cluster is the column intended for model conditioning; category is kept as the finer-grained original label.

Domain clusters

The 56 raw category tags were consolidated into 12 macro-clusters — many of the original tags were the same underlying motion/physics under a different label, and splitting them further than this leaves too little data per domain for a small model to learn each as distinct. Each cluster is capped at 4,000 clips after cutting, so no cluster dominates training signal.

Cluster	Raw categories folded in	Raw source videos
fluid_liquid	fluid_simulation, liquid_dynamics, fluid_obstacles, splashes_droplets, water_surface_events, submersion, buoyancy, waves_ocean_physics, ink_dye_simulation	2,699
optical_abstract	light_shadow_motion, reflection_motion, abstract_visuals, particles, microscopic, camera_motion, camera_observation, surface_interaction, complex_multi_object	2,696
human_motion	human_interaction, everyday_actions, sports_fitness, cooking, eating_drinking, hand_object_interaction, crowd_motion	2,444
granular_soft	sand_granular_motion, cloth_simulation, soft_body, melting_freezing, evaporation_condensation, snow_ice_motion	1,797
rigid_mechanical	rigid_body_physics, destruction, chain_reactions, collision_events, mechanical_motion, assembly_disassembly, opening_closing	1,657
nature_weather_time	nature_weather, landscapes, natural_cycles, growth_transformation, timelapse	1,493
fire_smoke	smoke_fire_particles, fire_heat, fire_spread, smoke_dynamics	1,200
space	nasa	1,003
animal_motion	animals, animal_interactions	804
vehicles_transport	vehicles, vehicle_interactions	600
urban_architecture	urban, architecture	599
underwater	underwater	300

Licensing

License is tracked per clip in the license column (both the raw and clip-level manifests), not as a single blanket license, because sources differ:

• Pexels / Pixabay: footage is under each platform's free-to-use content license at the time of collection. These generally permit free use including commercial projects, but don't guarantee indefinite terms or cover every edge case (e.g. implying endorsement, reselling unaltered stock footage). Check original_url for the specific item if you need current, authoritative terms.
• NASA: public domain.

No footage in this dataset was scraped from platforms that prohibit redistribution (e.g. YouTube); everything was sourced from CC0/permissively-licensed libraries or public-domain archives specifically to avoid that complication.

Intended use

Training data for a small (~150-200M parameter) from-scratch video diffusion model, operating on compressed latents from a frozen video tokenizer, conditioned on cluster identity. Not intended as a general-purpose, production-scale video dataset — clip count and per-domain diversity are sized for a small model's capacity, not for broad generalization at scale.

Limitations

• Clips are short (2s) and fixed-resolution/framerate by design — not suited for tasks needing longer temporal context or higher resolution without separate re-processing.
• Per-cluster clip counts are capped, so within-cluster diversity reflects a subsample of the raw footage, not the full raw pool.
• category labels were assigned from the original search query per video, not manually verified per clip — treat them as approximate, not ground truth.