Initial dataset card for 503-company longitudinal sample

README.md

@@ -1,3 +1,159 @@
----
-license: cc-by-nc-4.0
----
+---
+license: cc-by-nc-4.0
+task_categories:
+- tabular-classification
+- tabular-regression
+- time-series-forecasting
+- feature-extraction
+language:
+- en
+tags:
+- longitudinal
+- workforce
+- companies
+- historical
+- panel
+- training
+pretty_name: longitudinal_503companies_1950_2020
+size_categories:
+- 10K<n<100K
+---
+# Vivameda Longitudinal Company Sample (1950–2020)
+
+A 503-company longitudinal panel covering 70 years of company-level workforce evolution. Designed as a foundational training substrate for AI systems reasoning about organizations, companies, and human capital over time.
+
+**This is a sample drawn from the full Vivameda universe of 4.2M companies and 48M company-year observations.**
+
+## Why this dataset exists
+
+Most AI systems reasoning about companies suffer from severe recency bias. Their training data is scraped from the last few years of the web, which means they have only ever seen the post-2015 platform economy.
+
+This sample exposes models to seventy years of organizational evolution: postwar expansion, stagflation, the personal computer era, the dot-com boom and bust, the 2008 financial crisis, the platform era, and the late-cycle zero-interest-rate period. Each row describes a single company's observable state in a single year — its size, growth, workforce structure, capability mix, and computed signal flags.
+
+The dataset is fixed. It does not update. ML training runs against it produce reproducible results.
+
+## What's in the sample
+
+| | |
+|---|---|
+| Companies | 503 |
+| Company-year records | 25,988 |
+| Year range | 1950–2020 (71 years) |
+| Columns | 34 |
+| Schema layers | workforce, growth, tenure, role mix, capability mix, signal flags |
+
+The 503 companies are curated across five depth tiers to ensure the sample exercises every layer of the schema:
+
+| Tier | Companies | Description |
+|---|---|---|
+| 70-year anchor | 99 | Observable across the full 1950–2020 range. Large historical entities including IBM, Ford, Chevron, and major institutions. |
+| 60–69 year anchor | 147 | Postwar-era giants and long-running institutions. |
+| 50–59 year anchor | 147 | Late-20th-century emergent corporations. |
+| 40–49 year anchor | 99 | Tech and consumer companies founded in the 1970s–1980s. |
+| Modern showcase | 11 | Young high-profile companies including NVIDIA, SpaceX, Airbnb, Stripe, DoorDash, Palantir, Datadog, Databricks, Coinbase, Shopify, Figma. |
+
+## Coverage by era
+
+Not every row contains every layer. Coverage layers onto the base panel as source density increases through time. The `record_depth` column labels what is available in each row, so you can filter cleanly to the level of enrichment you need.
+
+| Era | Record depth | Rows |
+|---|---|---|
+| 1950–1989 | panel_only | 11,182 |
+| 1990–2009 | panel_plus_capability | 9,318 |
+| 2010–2016 | panel_plus_capability_plus_role | 3,476 |
+| 2017–2020 | full_record | 2,012 |
+
+NULL values in this dataset are meaningful. They reflect insufficient signal density at that point in time, not missing data points. Filter accordingly.
+
+## Schema
+
+### Identity
+- `company_id` — stable lowercase-dashed identifier, primary join key
+- `company_name` — human-readable display name
+- `company_domain` — primary web domain if known
+- `industry` — normalized taxonomy of 75 categories
+- `country` — country of corporate headquarters
+- `year` — observation year (1950–2020)
+- `sample_tier` — which curation tier the company was drawn from (sample-only, not present at universe scale)
+- `record_depth` — enrichment layers available in this row
+
+### Workforce and tenure
+- `headcount_observed` — observed employee count, directly observable not corporate-reported
+- `company_size_midpoint`, `company_size_bucket` — size aggregates
+- `net_headcount_change` — year-over-year headcount delta
+- `avg_tenure_years`, `tenure_bucket` — workforce tenure metrics
+
+### Growth
+- `growth_rate_yoy`, `prev_growth_rate_yoy` — year-over-year growth percentages
+- `growth_is_reliable` — passes internal plausibility checks
+- `growth_bucket` — categorical bucket from "Shrinking fast" to "Hyper growth"
+
+### Role mix (2010–2020)
+- `primary_role_bucket` — dominant role category from 9 canonical buckets
+- `primary_role_pct` — percentage of workforce in that role
+- `distinct_role_buckets` — count of distinct role categories present
+- `role_diversity_bucket`, `role_coverage_pct`
+
+### Capability mix (1990–2020)
+- `top_capability_1`, `top_capability_1_pct` (and ranks 2 and 3)
+- `capability_coverage_pct` — completeness indicator
+
+### Signal flags
+Four computed binary flags capturing notable year-over-year dynamics. Definitions are deterministic and reproducible from the raw growth fields.
+
+- `early_scaling_flag` — sustained fast growth at emerging scale
+- `contraction_flag` — transition from growth to meaningful decline
+- `recovery_signal_flag` — reversal from decline to growth
+- `growth_acceleration_flag` ��� increasing growth above a meaningful baseline
+
+ML practitioners will often prefer to compute their own signals from the raw growth fields, tailored to specific use cases. Raw fields are always available, including in rows where flags are NULL.
+
+## Use cases
+
+This sample is designed for:
+
+- Pre-training and fine-tuning models that reason about companies, workforces, or organizational evolution
+- Evaluating model behavior on out-of-distribution historical regimes
+- Time-series forecasting on long-horizon company panels
+- Feature engineering and embedding generation at the company-year grain
+- Behavioral pattern extraction across multi-decade economic regimes
+
+## Loading the data
+
+```python
+import pandas as pd
+
+df = pd.read_csv("vivameda_longitudinal_sample_503companies_1950_2020.csv")
+
+print(df.shape)             # (25988, 34)
+print(df['year'].min(), df['year'].max())  # 1950, 2020
+print(df['company_id'].nunique())          # 503
+
+# Filter to a specific era's record depth
+modern = df[df['record_depth'] == 'full_record']
+
+# Or to a specific tier
+historical = df[df['sample_tier'] == '70yr_anchor']
+```
+
+## About Vivameda
+
+Vivameda maintains a longitudinal universe of 4.2 million companies covering 1950–2020, with 48 million company-year records, role distributions, capability buckets, and tenure layers. The full universe powers organizational reasoning research, training data for AI systems, and predictive analytics for venture, credit, and strategy use cases.
+
+What hedge funds and quant funds have paid for over decades, Vivameda is making available to AI labs and researchers as training substrate.
+
+## Licensing
+
+This sample is released under **CC-BY-NC-4.0**. Free for research, academic, and non-commercial use with attribution. Commercial licensing of the full universe is available — contact Vivameda directly.
+
+## Contact
+
+- Web: [vivameda.com](https://vivameda.com)
+- Email: data@vivameda.com
+- Book a 20-minute call: [calendly.com/oli-nold/data-discussion-call](https://calendly.com/oli-nold/data-discussion-call)
+
+## Citation
+
+If you use this dataset in research, please cite:
+
+> Vivameda Longitudinal Company Sample (1950–2020), v1.0. Vivameda Ltd, April 2026.