Novelty AI Creative Generation Corpus

Description

Data Product Features

• Generation records — raw prompt, agent-enhanced prompt, model, provider, parameters, terminal status, error text, latency, reference-conditioning inputs, and parent_id lineage linking each retry to what it was derived from.
• Agent traces — 63,802 turns across 955 sessions, with structured tool-call events, token accounting, and latency on assistant turns.
• Orchestration runs — 2,704 runs with instructions, self-reported outcome, round counts, and parent/child delegation structure.
• Human verdicts — 927 client submissions resolving to approved / revision-requested / rejected, with 1,409 critique comments anchored to a specific second in the video.
• Media analysis — machine-generated shot boundaries, cut points, subject detection, and timestamped speech transcripts.
• Planning artifacts — 791 ordered storyboard scenes, scripts, and ad-copy variants tagged by marketing angle.

Distribution

• Data Volume:
  • 70,688 generation records, 42 columns — 66,748 completed (41,181 image / 24,773 video / 794 audio), 3,187 failed, 700 in-flight
  • 64,926 prompt-bearing assets — 61,049 completed, 3,180 failed
  • 63,802 agent messages (~1.39B tokens) · 2,704 runs · 955 sessions
  • 927 review versions · 1,409 review comments · 791 storyboard scenes · 178 code snapshots
  • 530 project briefs across 94 brand workspaces
  • Total: ~145,000 rows, ~66,700 media files

Usage

• Image/video post-training — paired prompt/output at production quality, plus a failure set for reward modeling.
• Prompt-rewriting models — ~61,000 aligned (intent → enhanced prompt → artifact) triples.
• Agentic tool-use training — 2,389 tool-call traces in a creative-production vertical.
• Preference and reward modeling — graded commercial approval signal with timestamped critique text.
• Self-correction research — parent_id chains reconstruct how professionals repair a bad generation.
• Cross-model evaluation — same brief generated across Gemini, Veo, Kling, Seedream, Seedance.

Coverage

• Geographic Coverage: English-language DTC creative, predominantly US and UK markets. Production team based in India.
• Time Range: 23 December 2025 – 6 August 2026.
• Demographics: No end-consumer data. 42 professional producer accounts; 94 brand workspaces spanning beauty, apparel, food and beverage, consumer electronics, and wellness.
• Providers represented: Gemini/Imagen (26,334), Kling (9,865), fal.ai (5,102), Veo (4,002), Seedream (1,894), Seedance (537), others (131). Agent layer: Claude Opus 4.6/4.7/4.8, Sonnet 4.6, Haiku 4.5.

License

Note: CC0 is not available for this corpus. It is an irrevocable public-domain dedication — it can only be granted by the rights holder over the entire work, and it contradicts the redistribution restrictions below, since CC0 material can be freely resold by anyone.

AI Training Rights

• Use the Data Product to train, fine-tune, and evaluate machine learning models, including large language models.
• Incorporate Data Product content into models and commercialize resulting model outputs.
• Create derivative works (model weights, embeddings, etc.) for any lawful purpose.

• The Data Product itself may not be sold, redistributed, or shared outside of licensed usage.
• Licensee must comply with all applicable laws, including data protection and privacy regulations.
• Licensee's rights are limited to those the Licensor can validly convey.

Who Can Use It

• Foundation model labs: Post-training for image/video generation and agentic tool use.
• Applied AI teams: Prompt-enhancement, auto-storyboarding, and shot-selection models.
• Data scientists: Reward and preference modeling on genuine commercial acceptance signal.
• Researchers: Human–agent iteration, self-correction, and failure recovery in a production loop.

Data Dictionary

| Column Name | Data Type | Description | Possible Values/Notes |
|-------------|-----------|-------------|-----------------------|
| asset_id | uuid | Generation attempt primary key | Stable across export |
| type | text | Media modality | image, video, audio |
| prompt | text | Operator-authored intent | Null on ~8% of rows |
| enhance_prompt | text | Agent-rewritten prompt dispatched | Core supervision target |
| model | varchar | Generation model | seedream-4-5, kling-v3-omni, veo, nano-banana-2 |
| provider | text | Serving provider | gemini, kling, fal, veo, seedream, seedance; null on 22,621 pre-instrumentation rows |
| status | varchar | Terminal state | completed, failed, generating |
| error_message | text | Provider failure reason | Populated on 3,187 rows |
| aspect_ratio / resolution | varchar | Output parameters | 9:16, 1:1, 16:9 / 1K–4K |
| images_used / videos_used | jsonb | Reference conditioning inputs | Array of asset refs |
| parent_id | uuid | Asset this was derived from | Null = root of iteration chain |
| analysis | jsonb | Shot boundaries, cut points, subjects | Raw-second timestamps; null = unanalyzed |
| transcript | jsonb | Timestamped speech | Video/audio with speech |
| generation_started_at / _completed_at | timestamptz | Dispatch and return | Yields per-request latency |
| is_favourite | boolean | Editor quality flag | Weak positive signal |
| message_role | varchar | Speaker in agent trace | user, assistant |
| message_content | varchar | Turn text | |
| tool_events | jsonb | Structured tool invocations | 2,389 assistant turns |
| input_tokens / output_tokens | integer | Per-turn accounting | |
| duration_ms / ttft_ms | integer | Latency, time-to-first-token | |
| run_instructions | text | Task given to agent | agent_runs |
| parent_run_id | uuid | Delegating parent | Multi-agent structure |
| review_status | text | Client verdict | pending_review, approved, revision_requested, rejected |
| comment_content | text | Reviewer critique | 1,409 rows |
| timestamp_sec | float8 | Point in video critiqued | Temporally grounded |
| scene_order / scene_description | integer / text | Storyboard position and shot intent | |

Additional Notes

1. Upstream provider terms. Every asset came from a third-party generation API and the agent layer from Anthropic models. Provider terms commonly restrict using outputs to train competing models. May require carving the corpus by provider or written permission.
2. Client work product. Prompts, briefs, and finished creative were produced under DTC client engagements; most such agreements assign work product to the client and impose confidentiality. Brand identity is recoverable from asset content, so pseudonymization doesn't resolve this.
3. Personal data. 42 producer profiles and reviewer names/emails on 1,409 comments require removal or a lawful basis before transfer.