Skip to content

data_card.md

Latest commit

 

History

History
121 lines (95 loc) · 4.97 KB

data_card.md

File metadata and controls

121 lines (95 loc) · 4.97 KB

Data Card — Oransim Synthetic Training Corpus

  • Dataset name: Oransim Synthetic Marketing Campaigns
  • Version: v1.0 (2026-04-18)
  • License: Apache-2.0 / CC0 (the data itself is public domain)
  • Companion model card: model_card.md

Summary

A fully synthetic corpus of simulated marketing campaigns produced by backend/scripts/gen_synthetic_data.py. No real KOL, note, user, or brand data enters this dataset. It is the exclusive training source for all pretrained models shipped with Oransim OSS.

Generation process

Artefact Process Distributional basis
KOL demographics IPF-calibrated sampling Published Chinese creator-economy aggregates
KOL fan counts Log-normal per niche Industry-reported niche-level creator tier distributions
KOL engagement rates Beta per niche Industry-reported engagement benchmarks
Notes (text + metrics) Template-filling + KOL linkage Aggregate note-per-day + engagement-rate priors
Fan profiles Static aggregate per niche Public demographic yearbooks
Training scenarios Budget × KOL × niche sampling + Hill + fatigue apply_budget_curves() + stochastic noise
Event streams Hawkes-like self-exciting PP Calibrated to match 14-day engagement envelope

Crucial property: no real instance-level data enters the pipeline. All samples are draws from distributions matched to published aggregate statistics, not from real observations.

Shipped files (v0.1.2-alpha)

File Rows Size Purpose
synthetic_kols.json 200 49 KB KOL pool
synthetic_notes.json 500 208 KB Note pool (English + Chinese text)
notes_v3.json 500 208 KB Back-compat alias for synthetic_notes.json
synthetic_fan_profiles.json 10 niches 4 KB Per-niche demographic distributions
niche_priors_calibrated.json 10 niches 4 KB Back-compat alias
scenarios_v0_1.jsonl 2 000 770 KB Training scenarios for the world model
event_streams_v0_1.jsonl 100 × 14-day streams 1.1 MB Training corpus for the Neural Hawkes

Total shipped demo data: 2.3 MB. The deterministic seed (42) ensures byte-identical regeneration. Users can scale up by rerunning the generator with their own --n-* arguments.

Canonical schema

All records conform to the Pydantic schemas in backend/oransim/data/schema/canonical.py:

  • CanonicalKOL — 200 instances
  • CanonicalNote — 500 instances
  • CanonicalFanProfile — 10 instances (one per niche)
  • CanonicalScenario — 2 000 instances
  • Event streams use the flat list-of-(time, event_type) form documented in oransim/diffusion/base.py.

Schema version: 1.1.

Intended use

Use case Supported?
Training the shipped LightGBM baseline ✅
Training the Causal Transformer world model ✅
Training the Causal Neural Hawkes diffusion model ✅
OrancBench reproducibility ✅
Testing platform-adapter code ✅
Benchmarking alternative causal models ✅
Real-world deployment ❌ — use a platform-specific DataProvider
Academic benchmarking vs real data ❌ — use held-out real data

Known limitations

  1. No real-instance fidelity — the corpus matches published aggregate statistics, not real individual campaigns. Metrics derived against this corpus do not transfer to production without re-calibration.
  2. Chinese-market bias — XHS-focused priors. Global-market deployment requires platform-specific DataProviders and recalibration.
  3. Text-only modality — no image/video creative embeddings. Multimodal generation lands in v0.5.
  4. No adversarial examples — fraud, bots, spam, and brand-safety edge cases are not explicitly modelled.
  5. Stationary distributions — no temporal drift is simulated. Real creator economies shift month-over-month.

Ethical considerations

  • The corpus contains no personal information about real users, KOLs, or brands. Synthetic names are drawn from a randomised vocabulary of English adjective-noun pairs.
  • OSS users training on this dataset can freely redistribute derived models under Apache-2.0.
  • OranAI Enterprise Edition trains on proprietary real-world data under separate agreements; those models are not distributed with the OSS release.

Reproducibility

python -m backend.scripts.gen_synthetic_data \
    --out data/synthetic --seed 42 \
    --n-kols 200 --n-notes 500 --n-scenarios 2000 --n-streams 100

Deterministic output. Identical seed → byte-identical files.

Versioning

Version Changes
v1.0 First documented release. 2.3 MB demo bundle shipped with v0.1.2-alpha.
v0.5 Planned: 500 k+ scenarios, comparative Copula vs GMM vs VAE generator study.
v1.0+ Planned: 5 M multimodal scenarios.

Maintenance

Please report data-quality issues via https://github.com/OranAi-Ltd/oransim/issues. Contact cto@orannai.com for commercial-data-licensing enquiries.