FlyRank internship data

Two datasets. The small one ships in this repo; the big one is hosted and gated.

1. Starter dataset (in this repo)

data/raw/content_refresh_anonymized.csv — 30,000 rows × 44 columns, one row per pseudonymized
content item, 32 clients, trailing-90-day metrics. Full column reference: docs/data-dictionary.md
(keep it open). The gotchas that cause 90% of mistakes:

• Rate columns are ×100 percentages: ctr = 0.76 means 0.76%, not 76%. Applies to ctr,
  engagement_rate, scroll_rate, ai_traffic_pct, trend_pct.
• avg_position = 0 means "no data", not rank zero (1,205 rows).
• scroll_rate and ai_traffic_pct can exceed 100 — numerator and denominator come from
  different measurement systems. Not a bug; read the dictionary.
• The label trap: is_declining_label is derived from trend_direction, which is computed
  from trend_pct. Therefore trend_direction and trend_pct are NEVER features.
• Missingness follows content_type (one type has ~100% missing keyword data; another ~28%
  missing word_count) — a blind fillna(0) injects a category signal. Add has_-flags instead.
• IDs (content_id, client_id) are pseudonyms: grouping/joining/splitting only, never
  features. Use client_id for grouped train/test splits.

2. Warehouse release (Hugging Face, gated — instant approval)

hf://datasets/FlyRank/internship-warehouse — build v20260703:

Table	Rows	Grain
dim_clients	104	one per pseudonymized client
dim_content	519,606	one per content item
fact_content_daily_performance	78,835,655	report_date × client × content (partitioned by month)
fact_content_daily_performance_sample	~11.7M	same grain — the FINAL MONTH (June 2026) only; fine for query mechanics, NEVER for label logic
fact_content_query_90d	2,414,248	client × content × query hash, fixed 90-day window

Dates: 2025-01-27 → 2026-06-30 (~17 months). Position column is gsc_avg_position.

Panel warnings (they are real):

• History depth differs wildly per client — check dim_clients.gsc_data_start before defining
  any time window; prefer per-client windows over one global calendar window.
• Rows before a client's ga4_data_start have GA4 columns zero-FILLED with
  ga4_data_available = FALSE — filter on the flag; zeros there are not "no engagement".
• A third of clients have little or no usable search/analytics history — expect to filter.
• The query table's per-content context columns repeat on every row: ANY_VALUE() them, never
  SUM(). Its 90-day window overlaps the snapshot's final months — if your label lives in the
  last 30 days, only *_prev30-style columns are safe features (window alignment first!).

Access + iteration rules: request gate access once (instant), READ token via Colab
Secrets (HF_TOKEN) or env var — never pasted in a cell (public repo!). Iterate on a
mid-panel month partition (e.g. month=2026-03), not the _sample: the _sample is the
panel's LAST month, i.e. the natural outcome window of any past→future label — develop label
logic there and you are developing inside your own test window. Treat the final month as a
sealed test month. Run the full 79M scan ONCE when the query is final and cache the result to
work/outputs/. Repeated full scans hit HTTP 429 rate limits.

How to verify

• Start any session with COUNT(*) + MIN/MAX(report_date) on your table and match the
  numbers above.
• Grain-probe before aggregating: GROUP BY <grain cols> HAVING COUNT(*) > 1 LIMIT 5 → empty.
• If a number looks absurd (a 400% rate, a giant ratio from 12 rows), re-read the gotchas
  before celebrating.