zhangyanye/forge-algr-tiny-sampled-1m-fullhist-all5core-eval10k

FORGE AL-GR Tiny Sampled 1M FullHist All-Data 5-Core Eval10k

This dataset is derived from the existing all-data 5-core FORGE export.

Filtering policy:

• Keep the existing sampled full-data iterative item 5-core result unchanged.
• Reassign train/valid/test split after 5-core.
• Use fixed evaluation sizes: 10,000 validation rows and 10,000 test rows.
• Put all remaining rows into train.
• Full user_history is retained; no recent-10 truncation is applied.

Statistics:

• records: 734,998
• train: 714,998
• valid: 10,000
• test: 10,000
• items: 2,359,297
• average history interactions: 49.12
• max observed history length after all-data 5-core: 100
• valid+test target cold ratio vs train history only: 0.000000%
• valid+test target cold ratio vs train history+target: 0.000000%

Generated files:

• pid2emb.parquet: pid plus float32[512] item embedding.

• pid2caption.parquet: pid plus title; titles filled from old mapping and forge/Item-Info/item_title.csv. Filled: 2,329,159; empty: 30,138.

• pid2sid.parquet: pid plus sid=list<int64>[3]; sid uses final_sid, falls back to base_sid, then [-1, -1, -1] if unavailable. Rows: 2,359,297; base fallback: 65; missing fallback: 5.